Anda di halaman 1dari 78

NEWS for R version 3.4.

1 (2017-06-30)


The deprecated support for PCRE versions older than 8.20 has been removed.

getParseData() gave incorrect column information when code contained multi-byte
characters. (PR#17254)
Asking for help using expressions like ?stats::cor() did not work. (PR#17250)
readRDS(url(....)) now works.
R CMD Sweave again returns status = 0 on successful completion.
Vignettes listed in .Rbuildignore were not being ignored properly. (PR#17246)
file.mtime() no longer returns NA on Windows when the file or directory is be-
ing used by another process. This affected installed.packages(), which is now
protected against this.
R CMD INSTALL Windows .zip file obeys --lock and --pkglock flags.
(Windows only) The choose.files() function could return incorrect results when
called with multi = FALSE. (PR#17270)
aggregate(<data.frame>, drop = FALSE) now also works in case of near-equal
numbers in by. (PR#16918)
fourfoldplot() could encounter integer overflow when calculating the odds ratio.
parse() no longer gives spurious warnings when extracting srcrefs from a file not
encoded in the current locale.
This was seen from R CMD check with inst/doc/*.R files, and check has some
additional protection for such files.
print.noquote(x) now always returns its argument x (invisibly).
Non-UTF-8 multibyte character sets were not handled properly in source references.


(Unix-alike) The default methods for download.file() and url() now choose
"libcurl" except for file:// URLs. There will be small changes in the format
and wording of messages, including in rare cases if an issue is a warning or an error.
For example, when HTTP re-direction occurs, some messages refer to the final URL
rather than the specified one.
Those who use proxies should check that their settings are compatible (see ?down-
load.file: the most commonly used forms work for both "internal" and
table() has been amended to be more internally consistent and become back com-
patible to R 2.7.2 again. Consequently, table(1:2, exclude = NULL) no longer
contains a zero count for <NA>, but useNA = "always" continues to do so.
summary.default() no longer rounds, but its print method does resulting in less
extraneous rounding, notably of numbers in the ten thousands.
factor(x, exclude = L) behaves more rationally when x or L are character vectors.
Further, exclude = <factor> now behaves as documented for long.
Arithmetic, logic (&, |) and comparison (aka relational, e.g., <, ==) operations with
arrays now behave consistently, notably for arrays of length zero.
Arithmetic between length-1 arrays and longer non-arrays had silently dropped the
array attributes and recycled. This now gives a warning and will signal an error in
the future, as it has always for logic and comparison operations in these cases (e.g.,
compare matrix(1,1) + 2:3 and matrix(1,1) < 2:3).
The JIT (Just In Time) byte-code compiler is now enabled by default at its level 3.
This means functions will be compiled on first or second use and top-level loops will
be compiled and then run. (Thanks to Tomas Kalibera for extensive work to make
this possible.)
For now, the compiler will not compile code containing explicit calls to browser():
this is to support single stepping from the browser() call.
JIT compilation can be disabled for the rest of the session using com-
piler::enableJIT(0) or by setting environment variable R_ENABLE_JIT to 0.
xtabs() works more consistently with NAs, also in its result no longer setting them
to 0. Further, a new logical option addNA allows to count NAs where appropriate.
Additionally, for the case sparse = TRUE, the results dimnames are identical to the
default cases.
Matrix products now consistently bypass BLAS when the inputs have NaN/Inf values.
Performance of the check of inputs has been improved. Performance when BLAS is
used is improved for matrix/vector and vector/matrix multiplication (DGEMV is now
used instead of DGEMM).
One can now choose from alternative matrix product implementations via op-
tions(matprod = ). The "internal" implementation is not optimized for speed
but consistent in precision with other summations in R (using long double accumu-
lators where available). "blas" calls BLAS directly for best speed, but usually with
undefined behavior for inputs with NaN/Inf.


User errors such as integrate(f, 0:1, 2) are now caught.

Add signature argument to debug(), debugonce(), undebug() and isdebugged()
for more conveniently debugging S3 and S4 methods. (Based on a patch by Gabe
Add utils::debugcall() and utils::undebugcall() for debugging the function
that would be called by evaluating the given expression. When the call is to an
S4 generic or standard S3 generic, debugcall() debugs the method that would be
dispatched. A number of internal utilities were added to support this, most notably
utils::isS3stdGeneric(). (Based on a patch by Gabe Becker.)
Add utils::strcapture(). Given a character vector and a regular expression con-
taining capture expressions, strcapture() will extract the captured tokens into a
tabular data structure, typically a data.frame.
str() and strOptions() get a new option drop.deparse.attr with improved but
changed default behaviour for expressions. For expression objects x, str(x) now
may remove extraneous white space and truncate long lines.
str(<looooooooong_string>) is no longer very slow; inspired by Mikko Korpelas
proposal in PR#16527.
str(x)s default method is more accurate and hence somewhat more generous in dis-
playing character vectors; this will occasionally change R outputs (and need changes
to some *.Rout(.save) files).
For a classed integer vector such as x <- xtabs(~ c(1,9,9,9)), str(x) now shows
both the class and "int", instead of only the latter.
isSymmetric(m) is much faster for large asymmetric matrices m via pre-tests and a
new option tol1 (with which strict back compatibility is possible but not the default).
The result of eigen() now is of class "eigen" in the default case when eigenvectors
are computed.
Zero-length date and date-time objects (of classes "POSIX[cl]?t") now print() rec-
xy.coords() and xyz.coords() get a new setLab option.
The method argument of sort.list(), order() and gains an "auto"
option (the default) which should behave the same as before when method was not
stopifnot(E, ..) now reports differences when E is a call to all.equal() and that
is not true.
boxplot(<formula>, *) gain optional arguments drop, sep, and lex.order to pass
to split.default() which itself gains an argument lex.order to pass to interac-
tion() for more flexibility.
The plot() method for ppr() has enhanced default labels (xmin and main). gains an explicit useHash option (with a back compatible default).
identical() gains an ignore.srcref option which drops "srcref" and similar at-
tributes when true (as by default).
diag(x, nrow = n) now preserves typeof(x), also for logical, integer and raw x
(and as previously for complex and numeric).
smooth.spline() now allows direct specification of lambda, gets a hatvalues()
method and keeps tol in the result, and optionally parts of the internal matrix

addNA() is faster now, e.g. when applied twice. (Part of PR#16895.)

New option rstandard(<lm>, type = "predicted") provides the PRESSrelated
leave-one-out cross-validation errors for linear models.
After seven years of deprecation, duplicated factor levels now produce a warning when
printed and an error in levels<- instead of a warning.
Invalid factors, e.g., with duplicated levels (invalid but constructable) now give a
warning when printed, via new function .valid.factor().
sessionInfo() has been updated for Apples change in OS naming as from 10.12
(macOS Sierra vs OS X El Capitan).
Its toLatex() method now includes the running component.
options(interrupt=) can be used to specify a default action for user interrupts. For
now, if this option is not set and the error option is set, then an unhandled user
interrupt invokes the error option. (This may be dropped in the future as interrupt
conditions are not error conditions.)
In most cases user interrupt handlers will be called with a "resume" restart available.
Handlers can invoke this restart to resume computation. At the browser prompt the
r command will invoke a "resume" restart if one is available. Some read operations
cannot be resumed properly when interrupted and do not provide a "resume" restart.
Radix sort is now chosen by method = "auto" for for double vectors
(and hence used for sort() for unclassed double vectors), excluding long vectors. = "radix") no longer rounds double vectors.
The default and data.frame methods for stack() preserve the names of empty
elements in the levels of the ind column of the return value. Set the new drop
argument to TRUE for the previous behavior.
Speedup in simplify2array() and hence sapply() and mapply() (for the case of
names and common length > 1), thanks to Suharto Anggonos PR#17118.
table(x, exclude = NULL) now sets useNA = "ifany" (instead of "always"). To-
gether with the bug fixes for this case, this recovers more consistent behaviour com-
patible to older versions of R. As a consequence, summary() for a logical vector no
longer reports (zero) counts for NA when there are no NAs.
dump.frames() gets a new option include.GlobalEnv which allows to also dump
the global environment, thanks to Andreas Kerstings proposal in PR#17116.
system.time() now uses message() instead of cat() when terminated early, such
that suppressMessages() has an effect; suggested by Ben Bolker.
citation() supports inst/CITATION files from package source trees, with lib.loc
pointing to the directory containing the package.
try() gains a new argument outFile with a default that can be modified via op-
tions(try.outFile = .), useful notably for Sweave.
The unexported low-level functions in package parallel for passing serialized R objects
to and from forked children now support long vectors on 64-bit platforms. This
removes some limits on higher-level functions such as mclapply() (but returning
gigabyte results from forked processes via serialization should be avoided if at all
Connections now print() without error even if invalid, e.g. after having been de-
apropos() and find(simple.words = FALSE) no longer match object names start-
ing with . which are known to be internal objects (such as .__S3MethodsTable__.).

Convenience function hasName() has been added; it is intended to replace the common
idiom !is.null(x$name) without the usually unintended partial name matching.
strcapture() no longer fixes column names nor coerces strings to factors (suggested
by Bill Dunlap).
strcapture() returns NA for non-matching values in x (suggested by Bill Dunlap).
source() gets new optional arguments, notably exprs; this is made use of in the new
utility function withAutoprint().
sys.source() gets a new toplevel.env argument. This argument is useful for frame-
works running package tests; contributed by Tomas Kalibera.
Sys.setFileTime() and file.copy( = TRUE) will set timestamps with
fractions of seconds on platforms/filesystems which support this.
(Windows only.) now returns file timestamps including fractions of
seconds; it has done so on other platforms since R 2.14.0. (NB: some filesystems do
not record modification and access timestamps to sub-second resolution.)
The license check enabled by options(checkPackageLicense = TRUE) is now done
when the packages namespace is first loaded.
ppr() and supsmu() get an optional trace argument, and
ppr(.., sm.method = ..spline) is no longer limited to sample size n 2500.
The POSIXct method for print() gets optional tz and usetz arguments, thanks to
a report from Jennifer S. Lyon.
New function check_packages_in_dir_details() in package tools for analyzing
package-check log files to obtain check details.
Package tools now exports function CRAN_package_db() for obtaining information
about current packages in the cran package repository, and several functions for
obtaining the check status of these packages.
The (default) Stangle driver Rtangle allows annotate to be a function and gets a
new drop.evalFALSE option.
The default method for quantile(x, prob) should now be monotone in prob, even
in border cases, see PR#16672. now tries to extract an email address from a BugReports field, and
if there is none, from a Contacts field.
The format() and print() methods for object.size() results get new options
standard and digits; notably, standard = "IEC" and standard = "SI" allow more
standard (but less common) abbreviations than the default ones, e.g. for kilobytes.
(From contributions by Henrik Bengtsson.)
If a reference class has a validity method, validObject will be called automatically
from the default initialization method for reference classes.
tapply() gets new option default = NA allowing to change the previously hardcoded
read.dcf() now consistently interprets any whitespace to be stripped to include
The maximum number of DLLs that can be loaded into R e.g. via dyn.load() can
now be increased by setting the environment variable R_MAX_NUM_DLLS before starting
Assigning to an element of a vector beyond the current length now over-allocates by a
small fraction. The new vector is marked internally as growable, and the true length

of the new vector is stored in the truelength field. This makes building up a vector
result by assigning to the next element beyond the current length more efficient,
though pre-allocating is still preferred. The implementation is subject to change and
not intended to be used in packages at this time.
Loading the parallel package namespace no longer sets or changes the .Random.seed,
even if R_PARALLEL_PORT is unset.
NB: This can break reproducibility of output, and did for a cran package.
Methods "wget" and "curl" for download.file() now give an R error rather than
a non-zero return value when the external command has a non-zero status.
Encoding name "utf8" is mapped to "UTF-8". Many implementations of iconv
accept "utf8", but not GNU libiconv (including the late 2016 version 1.15).
sessionInfo() shows the full paths to the library or executable files providing the
BLAS/LAPACK implementations currently in use (not available on Windows).
The binning algorithm used by bandwidth selectors bw.ucv(), bw.bcv() and bw.SJ()
switches to a version linear in the input size n for n > nb/2. (The calculations are
the same, but for larger n/nb it is worth doing the binning in advance.)
There is a new option PCRE_study which controls when grep(perl = TRUE) and
friends study the compiled pattern. Previously this was done for 11 or more input
strings: it now defaults to 10 or more (but most examples need many more for the
difference from studying to be noticeable).
grep(perl = TRUE) and friends can now make use of PCREs Just-In-Time mecha-
nism, for PCRE 8.20 on platforms where JIT is supported. It is used by default
whenever the pattern is studied (see the previous item). (Based on a patch from
Mikko Korpela.)
This is controlled by a new option PCRE_use_JIT.
Note that in general this makes little difference to the speed, and may take a little
longer: its benefits are most evident on strings of thousands of characters. As a side
effect it reduces the chances of C stack overflow in the PCRE library on very long
strings (millions of characters, but see next item).
Warning: segfaults were seen using PCRE with JIT enabled on 64-bit Sparc builds.
There is a new option PCRE_limit_recursion for grep(perl = TRUE) and friends to
set a recursion limit taking into account Rs estimate of the remaining C stack space
(or 10000 if that is not available). This reduces the chance of C stack overflow, but
because it is conservative may report a non-match (with a warning) in examples that
matched before. By default it is enabled if any input string has 1000 or more bytes.
getGraphicsEvent() now works on X11(type = "cairo") devices. Thanks to Fred-
erick Eaton (for reviving an earlier patch).
There is a new argument onIdle for getGraphicsEvent(), which allows an R function
to be run whenever there are no pending graphics events. This is currently only
supported on X11 devices. Thanks to Frederick Eaton.
The deriv() and similar functions now can compute derivatives of log1p(), sinpi()
and similar one-argument functions, thanks to a contribution by Jerry Lewis.
median() gains a formal ... argument, so methods with extra arguments can be
strwrap() reduces indent if it is more than half width rather than giving an error.
(Suggested by Bill Dunlap.)

When the condition code in if(.) or while(.) is not of length one, an error instead
of a warning may be triggered by setting an environment variable, see the help page.
Formatting and printing of bibliography entries (bibentry) is more flexible and better
documented. Apart from setting options(citation.bibtex.max = 99) you can also
use print(<citation>, bibtex=TRUE) (or format(..)) to get the BibTeX entries
in the case of more than one entry. This also affects citation(). Contributions to
enable style = "html+bibtex" are welcome.

Entry points R_MakeExternalPtrFn and R_ExternalPtrFn are now declared in
header Rinternals.h to facilitate creating and retrieving an R external pointer
from a C function pointer without ISO C warnings about the conversion of function
There was an exception for the native Solaris C++ compiler to the dropping (in
R 3.3.0) of legacy C++ headers from headers such as R.h and Rmath.h this
has now been removed. That compiler has strict C++98 compliance hence does
not include extensions in its (non-legacy) C++ headers: some packages will need to
request C++11 or replace non-C++98 calls such as lgamma: see 1.6.4 of Writing R
Because it is needed by about 70 CRAN packages, headers R.h and Rmath.h still
use namespace std;
when included on Solaris.
When included from C++, the R headers now use forms such as std::FILE directly
rather than including the line
using std::FILE;
C++ code including these headers might be relying on the latter.
Headers R_ext/BLAS.h and R_ext/Lapack.h have many improved declarations
including const for double-precision complex routines. Inter alia this avoids warnings
when passing string literal arguments from C++11 code.
Headers for Unix-only facilities R_ext/GetX11Image.h, R_ext/QuartzDevice.h
and R_ext/eventloop.h are no longer installed on Windows.
No-longer-installed headers GraphicsBase.h, RGraphics.h, Rmodules/RX11.h
and Rmodules/Rlapack.h which had a LGPL license no longer do so.
HAVE_UINTPTR_T is now defined where appropriate by Rconfig.h so that it can be
included before Rinterface.h when CSTACK_DEFNS is defined and a C compiler (not
C++) is in use. Rinterface.h now includes C header stdint.h or C++11 header
cstdint where needed.
Package tools has a new function package_native_routine_registration_skeleton()
to assist adding native-symbol registration to a package. See its help and 5.4.1 of
Writing R Extensions for how to use it. (At the time it was added it successfully
automated adding registration to over 90% of cran packages which lacked it. Many
of the failures were newly-detected bugs in the packages, e.g. 50 packages called entry
points with varying numbers of arguments and 65 packages called entry points not
in the package.)


readline headers (and not just the library) are required unless configuring with
configure now adds a compiler switch for C++11 code, even if the compiler supports
C++11 by default. (This ensures that g++ 6.x uses C++11 mode and not its default
mode of C++14 with GNU extensions.)
The tests for C++11 compliance are now much more comprehensive. For gcc < 4.8,
the tests from R 3.3.0 are used in order to maintain the same behaviour on Linux
distributions with long-term support.
An alternative compiler for C++11 is now specified with CXX11, not CXX1X.
Likewise C++11 flags are specified with CXX11FLAGS and the standard (e.g.,
-std=gnu++11 is specified with CXX11STD.
configure now tests for a C++14-compliant compiler by testing some basic features.
This by default tries flags for the compiler specified by CXX11, but an alternative
compiler, options and standard can be specified by variables CXX14, CXX14FLAGS
and CXX14STD (e.g., -std=gnu++14).
There is a new macro CXXSTD to help specify the standard for C++ code,
e.g. -std=c++98. This makes it easier to work with compilers which default to
a later standard: for example, with CXX=g++6 CXXSTD=-std=c++98 configure will
select commands for g++ 6.x which conform to C++11 and C++14 where specified
but otherwise use C++98.
Support for the defunct IRIX and OSF/1 OSes and Alpha CPU has been removed.
configure checks that the compiler specified by $CXX $CXXFLAGS is able to compile
C++ code.
configure checks for the required header sys/select.h (or sys/time.h on legacy
systems) and system call select and aborts if they are not found.
If available, the POSIX 2008 system call utimensat will be used by
Sys.setFileTime() and file.copy( = TRUE). This may result in slightly
more accurate file times. (It is available on Linux and FreeBSD but not macOS.)
The minimum version requirement for libcurl has been reduced to 7.22.0, although
at least 7.28.0 is preferred and earlier versions are little tested. (This is to support
Debian 7 Wheezy LTS and Ubuntu Precise 12.04 LTS, although the latter is close
to end-of-life.)
configure tests for a C++17-compliant compiler. The tests are experimental and
subject to change in the future.

(Windows only) Tcl/Tk version 8.6.4 is now included in the binary builds. The
tcltk*.chm help file is no longer included; please consult the online help at http:
// instead.
The version of LAPACK included in the sources has been updated to 3.7.0: no new
routines have been added to R.

There is support for compiling C++14 or C++17 code in packages on suitable plat-
forms: see Writing R Extensions for how to request this.
The order of flags when LinkingTo other packages has been changed so their include
directories come earlier, before those specified in CPPFLAGS. This will only have an

effect if non-system include directories are included with -I flags in CPPFLAGS (and so
not the default -I/usr/local/include which is treated as a system include directory
on most platforms).
Packages which register native routines for .C or .Fortran need to be re-installed for
this version (unless installed with R-devel SVN revision r72375 or later).
Make variables with names containing CXX1X are deprecated in favour of those us-
ing CXX11, but for the time being are still made available via file etc/Makeconf.
Packages using them should be converted to the new forms and made dependent on
R (>= 3.4.0).

Running R CMD check --as-cran with _R_CHECK_CRAN_INCOMING_REMOTE_ false
now skips tests that require remote access. The remaining (local) tests typically
run quickly compared to the remote tests.
R CMD build will now give priority to vignettes produced from files in the vignettes
directory over those in the inst/doc directory, with a warning that the latter are
being ignored.
R CMD config gains a --all option for printing names and values of all basic con-
figure variables.
It now knows about all the variables used for the C++98, C++11 and C++14 stan-
R CMD check now checks that output files in inst/doc are newer than the source
files in vignettes.
For consistency with other package subdirectories, files named *.r in the tests
directory are now recognized as tests by R CMD check. (Wish of PR#17143.)
R CMD build and R CMD check now use the union of R_LIBS and .libPaths(). They
may not be equivalent, e.g., when the latter is determined by R_PROFILE.
R CMD build now preserves dates when it copies files in preparing the tarball. (Pre-
viously on Windows it changed the dates on all files; on Unix, it changed some dates
when installing vignettes.)
The new option R CMD check --no-stop-on-test-error allows running the remain-
ing tests (under tests/) even if one gave an error.
Check customization via environment variables to detect side effects of .Call() and
.External() calls which alter their arguments is described in 8 of the R Internals
R CMD check now checks any BugReports field to be non-empty and a suitable single
R CMD check --as-cran now NOTEs if the package does not register its native rou-
tines or does not declare its intentions on (native) symbol search. (This will become
a WARNING in due course.)


(Windows only) Function setInternet2() is defunct.
Installation support for readline emulations based on editline (aka libedit) is
Use of the C/C++ macro NO_C_HEADERS is defunct and silently ignored.

unix.time(), a traditional synonym for system.time(), has been deprecated.

structure(NULL, ..) is now deprecated as you cannot set attributes on NULL.
Header Rconfig.h no longer defines SUPPORT_OPENMP; instead use _OPENMP (as
documented for a long time).
(C-level Native routine registration.) The deprecated styles member of the
R_CMethodDef and R_FortranMethodDef structures has been removed. Packages us-
ing these will need to be re-installed for R 3.4.0.
The deprecated support for PCRE versions older than 8.20 will be removed in R 3.4.1.
(Versions 8.208.31 will still be accepted but remain deprecated.)

Getting or setting body() or formals() on non-functions for now signals a warning
and may become an error for setting.
match(x, t), duplicated(x) and unique(x) work as documented for complex num-
bers with NAs or NaNs, where all those containing NA do match, whereas in the case of
NaNs both real and imaginary parts must match, compatibly with how print() and
format() work for complex numbers.
deparse(<complex>, options = "digits17") prints more nicely now, mostly
thanks to a suggestion by Richie Cotton.
Rotated symbols in plotmath expressions are now positioned correctly on
x11(type = "Xlib"). (PR#16948)
as<-() avoids an infinite loop when a virtual class is interposed between a subclass
and an actual superclass.
Fix level propagation in unlist() when the list contains zero-length lists or factors.
Fix S3 dispatch on S4 objects when the methods package is not attached.
Internal S4 dispatch sets .Generic in the method frame for consistency with
standardGeneric(). (PR#16929)
Fix order(x, decreasing = TRUE) when x is an integer vector containing MAX_INT.
Ported from a fix Matt Dowle made to data.table.
Fix caching by callNextMethod(), resolves PR#16973 and PR#16974.
grouping() puts NAs last, to be consistent with the default behavior of order().
Point mass limit cases: qpois(-2, 0) now gives NaN with a warning and qgeom(1, 1)
is 0. (PR#16972)
table() no longer drops an "NaN" factor level, and better obeys exclude = <chr>,
thanks to Suharto Anggonos patch for PR#16936. Also, in the case of
exclude = NULL and NAs, these are tabulated correctly (again).
Further, table(1:2, exclude = 1, useNA = "ifany") no longer erroneously re-
ports <NA> counts.
Additionally, all cases of empty exclude are equivalent, and useNA is not overwritten
when specified (as it was by exclude = NULL).
wilcox.test(x, no longer errors out in cases where the confidence
interval is not available, such as for x = 0:2.
droplevels(f) now keeps <NA> levels when present.
In integer arithmetic, NULL is now treated as integer(0) whereas it was previously
treated as double(0).

The radix sort considers NA_real_ and NaN to be equivalent in rank (like the other
sort algorithms).
When index.return=TRUE is passed to, the radix sort treats NAs like
sort.list() does (like the other sort algorithms).
When in tabulate(bin, nbin) length(bin) is larger than the maximal integer, the
result is now of type double and hence no longer silently overflows to wrong values.
as.character.factor() respects S4 inheritance when checking the type of its argu-
ment. (PR#17141)
The factor method for print() no longer sets the class of the factor to NULL, which
would violate a basic constraint of an S4 object.
formatC(x, flag = f) allows two new flags, and signals an error for invalid flags
also in the case of character formatting.
Reading from file("stdin") now also closes the connection and hence no longer leaks
memory when reading from a full pipe, thanks to Gabor Csardi, see thread starting
Failure to create file in tempdir() for compressed pdf() graphics device no longer
errors (then later segfaults). There is now a warning instead of error and compression
is turned off for the device. Thanks to Alec Wysoker (PR#17191).
Asking for methods() on "|" returns only S3 methods. See
dev.capture() using Quartz Cocoa device (macOS) returned invalid components if
the back-end chose to use ARGB instead of RGBA image format. (Reported by Noam
seq("2", "5") now works too, equivalently to "2":"5" and = 1, by = 1) is now correct, other cases are integer (instead of double)
when seq() is integer too, and the non-finite error messages are consistent between
seq.default() and, no longer mentioning NaN etc.
rep(x, times) and, times) now work when times is larger than the
largest value representable in an integer vector. (PR#16932)
download.file(method = "libcurl") does not check for URL existence before at-
tempting downloads; this is more robust to servers that do not support HEAD or
range-based retrieval, but may create empty or incomplete files for aborted download
Bandwidth selectors bw.ucv(), bw.bcv() and bw.SJ() now avoid integer overflow for
large sample sizes.
str() no longer shows "list output truncated", in cases that list was not shown
at all. Thanks to Neal Fultz (PR#17219)
Fix for cairo_pdf() (and svg() and cairo_ps()) when replaying a saved display
list that contains a mix of grid and graphics output. (Report by Yihui Xie.)
The str() and as.hclust() methods for "dendrogram" now also work for deeply
nested dendrograms thanks to non-recursive implementations by Bradley Broom.
sample() now uses two uniforms for added precision when the uniform generator
is Knuth-TAOCP, Knuth-TAOCP-2002, or a user-defined generator and the population
size is 225 or greater.

If a vignette in the vignettes directory is listed in .Rbuildignore, R CMD build

would not include it in the tarball, but would include it in the vignette database,
leading to a check warning. (PR#17246)
tools::latexToUtf8() infinite looped on certain inputs. (PR#17138)
terms.formula() ignored argument names when determining whether two terms were
identical. (PR#17235)
callNextMethod() was broken when called from a method that augments the formal
arguments of a primitive generic.
Coercion of an S4 object to a vector during sub-assignment into a vector failed to
dispatch through the as.vector() generic (often leading to a segfault).
Fix problems in command completion: Crash (PR#17222) and junk display in Win-
dows, handling special characters in filenames on all systems.

Changes when redirection of a http:// URL to a https:// URL is encountered:
The internal methods of download.file() and url() now report that they can-
not follow this (rather than failing silently).
(Unix-alike) download.file(method = "auto") (the default) re-tries with
method = "libcurl".
(Unix-alike) url(method = "default") with an explicit open argument re-tries
with method = "libcurl". This covers many of the usages, e.g. readLines()
with a URL argument.

The configure check for the zlib version is now robust to versions longer than 5
characters, including 1.2.11.

Environmental variable _R_CHECK_TESTS_NLINES_ controls how R CMD check reports
failing tests (see 8 of the R Internals manual).


(C-level Native routine registration.) The undocumented styles field of the compo-
nents of R_CMethodDef and R_FortranMethodDef is deprecated.

vapply(x, *) now works with long vectors x. (PR#17174)
isS3method("") and similar are correct now. (PR#17171)
grepRaw(<long>, <short>, fixed = TRUE) now works, thanks to a patch by Mikko
Korpela. (PR#17132)
Package installation into a library where the package exists via symbolic link now
should work wherever Sys.readlink() works, resolving PR#16725.
"Cincinnati" was missing an "n" in the precip dataset.

Fix buffer overflow vulnerability in pdf() when loading an encoding file. Reported
by Talos (TALOS-2016-0227).
getDLLRegisteredRoutines() now produces its warning correctly when multiple
DLLs match, thanks to Matt Dowles PR#17184.
Sys.timezone() now returns non-NA also on platforms such as
Ubuntu 14.04.5 LTS, thanks to Mikko Korpelas PR#17186.
format(x) for an illegal "POSIXlt" object x no longer segfaults.
methods(f) now also works for f "(" or "{".
(Windows only) dir.create() did not check the length of the path to create, and so
could overflow a buffer and crash R. (PR#17206)
On some systems, very small hexadecimal numbers in hex notation would underflow
to zero. (PR#17199)
pmin() and pmax() now work again for ordered factors and 0-length S3 classed
objects, thanks to Suharto Anggonos PR#17195 and PR#17200. did not do any validity checking on a packages BugReports field. It
now ignores an empty field, removes leading whitespace and only attempts to open
http:// and https:// URLs, falling back to emailing the maintainer.
Bandwidth selectors bw.ucv() and bw.SJ() gave incorrect answers or incorrectly
reported an error (because of integer overflow) for inputs longer than 46341. Similarly
for bw.bcv() at length 5793.
Another possible integer overflow is checked and may result in an error report (rather
than an incorrect result) for much longer inputs (millions for a smooth distribution).
findMethod() failed if the active signature had expanded beyond what a particular
package used. (Example with packages XR and XRJulia on cran.)
qbeta() underflowed too early in some very asymmetric cases. (PR#17178)
R CMD Rd2pdf had problems with packages with non-ASCII titles in .Rd files (usu-
ally the titles were omitted).

extSoftVersion() now reports the version (if any) of the readline library in use.
The version of LAPACK included in the sources has been updated to 3.6.1, a bug-fix
release including a speedup for the non-symmetric case of eigen().
Use options(deparse.max.lines=) to limit the number of lines recorded in
.Traceback and other deparsing activities.
format(<AsIs>) looks more regular, also for non-character atomic matrices.
abbreviate() gains an option named = TRUE.
The online documentation for package methods is extensively rewritten. The goals
are to simplify documentation for basic use, to note old features not recommended
and to correct out-of-date information.
Calls to setMethod() no longer print a message when creating a generic function in
those cases where that is natural: S3 generics and primitives.



Versions of the readline library >= 6.3 had been changed so that terminal win-
dow resizes were not signalled to readline: code has been added using a explicit
signal handler to work around that (when R is compiled against readline >= 6.3).
configure works better with Oracle Developer Studio 12.5.

R CMD check reports more dubious flags in files src/Makevars[.in], including -w
and -g.
R CMD check has been set up to filter important warnings from recent versions of
gfortran with -Wall -pedantic: this now reports non-portable GNU extensions
such as out-of-order declarations.
R CMD config works better with paths containing spaces, even those of home direc-
tories (as reported by Ken Beath).


Use of the C/C++ macro NO_C_HEADERS is deprecated (no C headers are included
by R headers from C++ as from R 3.3.0, so it should no longer be needed).

The check for non-portable flags in R CMD check could be stymied by src/Makevars
files which contained targets.
(Windows only) When using certain desktop themes in Windows 7 or higher, Alt-
Tab could cause Rterm to stop accepting input. (PR#14406; patch submitted by Jan
pretty(d, ..) behaves better for date-time d (PR#16923).
When an S4 class name matches multiple classes in the S4 cache, perform a dynamic
search in order to obey namespace imports. This should eliminate annoying mes-
sages about multiple hits in the class cache. Also, pass along the package from the
ClassExtends object when looking up superclasses in the cache.
sample(NA_real_) now works.
Packages using non-ASCII encodings in their code did not install data properly on
systems using different encodings.
merge(df1, df2) now also works for data frames with column names "na.last",
"decreasing", or "method". (PR#17119)
contour() caused a segfault if the labels argument had length zero. (Reported by
Bill Dunlap.)
unique(warnings()) works more correctly, thanks to a new
duplicated.warnings() method.
findInterval(x, vec = numeric(), all.inside = TRUE) now returns 0s as doc-
umented. (Reported by Bill Dunlap.)
(Windows only) R CMD SHLIB failed when a symbol in the resulting library had the
same name as a keyword in the .def file. (PR#17130)
pmax() and pmin() now work with (more ?) classed objects, such as "Matrix" from
the Matrix package, as documented for a long time.

axis(side, x = D) and hence Axis() and plot() now work correctly for "Date"
and time objects D, even when time goes backward, e.g., with decreasing xlim.
(Reported by William May.)
str(I(matrix(..))) now looks as always intended.
plot.ts(), the plot() method for time series, now respects cex, lwd and lty. (Re-
ported by Greg Werbin.)
parallel::mccollect() now returns a named list (as documented) when called with
wait = FALSE. (Reported by Michel Lang.)
If a package added a class to a class union in another package, loading the first package
gave erroneous warnings about undefined subclass.
c()s argument use.names is documented now, as belonging to the (C internal) de-
fault method. In parallel, argument recursive is also moved from the generic to
the default method, such that the formal argument list of base generic c() is just
rbeta(4, NA) and similarly rgamma() and rnbinom() now return NaNs with a warn-
ing, as other r<dist>(), and as documented. (PR#17155)
Using options(checkPackageLicense = TRUE) no longer requires acceptance of the
licence for non-default standard packages such as compiler. (Reported by Mikko
split(<very_long>, *) now works even when the split off parts are long.
min() and max() now also work correctly when the argument list starts with
character(0). (PR#17160)
Subsetting very large matrices (prod(dim(.)) >= 2^31) now works thanks to
Michael Schubmehls PR#17158.
bartlett.test() used residual sums of squares instead of variances, when the argu-
ment was a list of lm objects. (Reported by Jens Ledet Jensen).
plot(<lm>, which = *) now correctly labels the contour lines for the standardized
residuals for which = 6. It also takes the correct p in case of singularities (also for
which = 5). (PR#17161)
xtabs(~ exclude) no longer fails from wrong scope, thanks to Suharto Anggonos
Reference class calls to methods() did not re-analyse previously defined methods,
meaning that calls to methods defined later would fail. (Reported by Charles Tilford).
findInterval(x, vec, = TRUE) misbehaved in some cases. (Reported
by Dmitriy Chernykh.)

R CMD INSTALL and hence install.packages() gave an internal error installing a
package called description from a tarball on a case-insensitive file system.
match(x, t) (and hence x %in% t) failed when x was of length one, and either
character and x and t only differed in their Encoding or when x and t where
complex with NAs or NaNs. (PR#16885.)

unloadNamespace(ns) also works again when ns is a namespace, as from

rgamma(1, Inf) or rgamma(1, 0, 0) no longer give NaN but the correct limit.
length(baseenv()) is correct now.
pretty(d, ..) for date-time d rarely failed when "halfmonth" time steps were tried
(PR#16923) and on inaccurate platforms such as 32-bit Windows or a configuration
with --disable-long-double; see comment #15 of PR#16761.
In text.default(x, y, labels), the rarely(?) used default for labels is now cor-
rect also for the case of a 2-column matrix x and missing y.
as.factor(c(a = 1L)) preserves names() again as in R < 3.1.0.
strtrim(""[0], 0[0]) now works.
Use of Ctrl-C to terminate a reverse incremental search started by Ctrl-R in the
readline-based Unix terminal interface is now supported when R was compiled
against readline >= 6.0 (Ctrl-G always worked). (PR#16603)
diff(<difftime>) now keeps the "units" attribute, as subtraction already did,

nchar(x, *)s argument keepNA governing how the result for NAs in x is deter-
mined, gets a new default keepNA = NA which returns NA where x is NA, except for
type = "width" which still returns 2, the formatting / printing width of NA.
All builds have support for https: URLs in the default methods for
download.file(), url() and code making use of them.
Unfortunately that cannot guarantee that any particular https: URL can be ac-
cessed. For example, server and client have to successfully negotiate a cryptographic
protocol (TLS/SSL, . . . ) and the servers identity has to be verifiable via the available
certificates. Different access methods may allow different protocols or use private cer-
tificate bundles: we encountered a https: CRAN mirror which could be accessed by
one browser but not by another nor by download.file() on the same Linux machine.

The print method for methods() gains a byclass argument.
New functions validEnc() and validUTF8() to give access to the validity checks for
inputs used by grep() and friends.
Experimental new functionality for S3 method checking, notably isS3method().
Also, the names of the R language elements are exported as character vector
str(x) now displays "Time-Series" also for matrix (multivariate) time-series, i.e.
when is.ts(x) is true.
(Windows only) The GUI menu item to install local packages now accepts *.tar.gz
files as well as *.zip files (but defaults to the latter).
New programmeRs utility function chkDots().

D() now signals an error when given invalid input, rather than silently returning NA.
(Request of John Nash.)
formula objects are slightly more first class: e.g., formula() or
new("formula", y ~ x) are now valid. Similarly, for "table", "ordered"
and "summary.table". Packages defining S4 classes with the above S3/S4 classes as
slots should be reinstalled.
New function strrep() for repeating the elements of a character vector.
rapply() preserves attributes on the list when how = "replace".
New S3 generic function sigma() with methods for extracting the estimated standard
deviation aka residual standard deviation from a fitted model.
news() now displays R and package news files within the HTML help system if it is
available. If no news file is found, a visible NULL is returned to the console.
as.raster(x) now also accepts raw arrays x assuming values in 0:255.
Subscripting of matrix/array objects of type "expression" is now supported.
type.convert("i") now returns a factor instead of a complex value with zero real
part and missing imaginary part.
Graphics devices cairo_pdf() and cairo_ps() now allow non-default values of the
cairographics fallback resolution to be set.
This now defaults to 300 on all platforms: that is the default documented by cairo-
graphics, but apparently was not used by all system installations.
file() gains an explicit method argument rather than implicitly using
getOption("url.method", "default").
Thanks to a patch from Tomas Kalibera, x[x != 0] is now typically faster than
x[which(x != 0)] (in the case where x has no NAs, the two are equivalent).
read.table() now always uses the names for a named colClasses argument (pre-
viously names were only used when colClasses was too short). (In part, wish of
(Windows only) download.file() with default method = "auto" and a ftps://
URL chooses "libcurl" if that is available.
The out-of-the box Bioconductor mirror has been changed to one using https://:
use chooseBioCmirror() to choose a http:// mirror if required.
The data frame and formula methods for aggregate() gain a drop argument.
available.packages() gains a repos argument.
The undocumented switching of methods for url() on https: and ftps: URLs
is confined to method = "default" (and documented).
smoothScatter() gains a ret.selection argument.
qr() no longer has a ... argument to pass additional arguments to methods.
[ has a method for class "table".
It is now possible (again) to replayPlot() a display list snapshot that was created
by recordPlot() in a different R session.
It is still not a good idea to use snapshots as a persistent storage format for R plots,
but it is now not completely silly to use a snapshot as a format for transferring an R
plot between two R sessions.
The underlying changes mean that packages providing graphics devices (e.g., Cairo,
RSvgDevice, cairoDevice, tikzDevice) will need to be reinstalled.

Code for restoring snapshots was contributed by Jeroen Ooms and JJ Allaire.
Some testing code is available at
tools::undoc(dir = D) and codoc(dir = D) now also work when D is a directory
whose normalizePath()ed version does not end in the package name, e.g. from a
abbreviate() has more support for multi-byte character sets it no longer removes
bytes within characters and knows about Latin vowels with accents. It is still only
really suitable for (most) European languages, and still warns on non-ASCII input.
abbreviate(use.classes = FALSE) is now implemented, and that is more suitable
for non-European languages.
match(x, table) is faster (sometimes by an order of magnitude) when x is of length
one and incomparables is unchanged, thanks to Peter Haverty (PR#16491).
More consistent, partly not back-compatible behavior of NA and NaN coercion to com-
plex numbers, operations less often resulting in complex NA (NA_complex_).
lengths() considers methods for length and [[ on x, so it should work automatically
on any objects for which appropriate methods on those generics are defined.
The logic for selecting the default screen device on OS X has been simplified: it is
now quartz() if that is available even if environment variable DISPLAY has been set
by the user.
The choice can easily be overridden via environment variable R_INTERACTIVE_DEVICE.
On Unix-like platforms which support the getline C library function,
system(*,intern = TRUE) no longer truncates (output) lines longer than 8192 char-
acters, thanks to Karl Millar. (PR#16544)
rank() gains a ties.method = "last" option, for convenience (and symmetry).
regmatches(invert = NA) can now be used to extract both non-matched and
matched substrings.
data.frame() gains argument fix.empty.names; gets new
cut.names, col.names and fix.empty.names.
plot(x ~ x, *) now warns that it is the same as plot(x ~ 1, *).
recordPlot() has new arguments load and attach to allow package names to be
stored as part of a recorded plot. replayPlot() has new argument reloadPkgs to
load/attach any package names that were stored as part of a recorded plot.
S4 dispatch works within calls to .Internal(). This means explicit S4 generics are
no longer needed for unlist() and as.vector().
Only font family names starting with "Hershey" (and not "Her" as before) are
given special treatment by the graphics engine.
S4 values are automatically coerced to vector (via as.vector) when subassigned into
atomic vectors.
findInterval() gets a option.
The version of LAPACK included in the sources has been updated to 3.6.0, including
those deprecated routines which were previously included. Ca 40 double-complex
routines have been added at the request of a package maintainer.
As before, the details of what is included are in src/modules/lapack/README and
this now gives information on earlier additions.
tapply() has been made considerably more efficient without changing functionality,
thanks to proposals from Peter Haverty and Suharto Anggono. (PR#16640)

match.arg(arg) (the one-argument case) is faster; so is (PR#16640)

The format method for object_size objects now also accepts binary units such as
"KiB" and e.g., "Tb". (Partly from PR#16649.)
Profiling now records calls of the form foo::bar and some similar cases directly rather
than as calls to <Anonymous>. Contributed by Winston Chang.
New string utilities startsWith(x, prefix) and endsWith(x, suffix). Also pro-
vide speedups for some grepl("^...",*) uses (related to proposals in PR#16490).
Reference class finalizers run at exit, as well as on garbage collection.
Avoid parallel dependency on stats for port choice and random number seeds.
The radix sort algorithm and implementation from data.table (forder) replaces the
previous radix (counting) sort and adds a new method for order(). Contributed by
Matt Dowle and Arun Srinivasan, the new algorithm supports logical, integer (even
with large values), real, and character vectors. It outperforms all other methods, but
there are some caveats (see ?sort).
The order() function gains a method argument for choosing between "shell" and
New function grouping() returns a permutation that stably rearranges data so that
identical values are adjacent. The return value includes extra partitioning information
on the groups. The implementation came included with the new radix sort.
rhyper(nn, m, n, k) no longer returns NA when one of the three parameters exceeds
the maximal integer.
switch() now warns when no alternatives are provided.
parallel::detectCores() now has default logical = TRUE on all platforms as
this was the default on Windows, this change only affects Sparc Solaris.
Option logical = FALSE is now supported on Linux and recent versions of OS X
(for the latter, thanks to a suggestion of Kyaw Sint).
hist() for "Date" or "POSIXt" objects would sometimes give misleading labels on
the breaks, as they were set to the day before the start of the period being displayed.
The display format has been changed, and the shift of the start day has been made
conditional on right = TRUE (the default). (PR#16679)
R now uses a new version of the logo (donated to the R Foundation by RStudio).
It is defined in .svg format, so will resize without unnecessary degradation when
displayed on HTML pagesthere is also a vector PDF version. Thanks to Dirk
Eddelbuettel for producing the corresponding X11 icon.
New function .traceback() returns the stack trace which traceback() prints.
lengths() dispatches internally.
dotchart() gains a pt.cex argument to control the size of points separately from
the size of plot labels. Thanks to Michael Friendly and Milan Bouchet-Valat for ideas
and patches.
as.roman(ch) now correctly deals with more diverse character vectors ch; also arith-
metic with the resulting roman numbers works in more cases. (PR#16779)
prcomp() gains a new option rank. allowing to directly aim for less than min(n,p)
PCs. The summary() and its print() method have been amended, notably for this

gzcon() gains a new option text, which marks the connection as text-oriented (so
e.g. pushBack() works). It is still always opened in binary mode.
The import() namespace directive now accepts an argument except which names
symbols to exclude from the imports. The except expression should evaluate to a
character vector (after substituting symbols for strings). See Writing R Extensions.
New convenience function Rcmd() in package tools for invoking R CMD tools from
within R.
New functions makevars_user() and makevars_site() in package tools to determine
the location of the user and site specific Makevars files for customizing package

R CMD check has a new option --ignore-vignettes for use with non-Sweave vi-
gnettes whose VignetteBuilder package is not available.
R CMD check now by default checks code usage (via codetools) with only the base
package attached. Functions from default packages other than base which are used in
the package code but not imported are reported as undefined globals, with a suggested
addition to the NAMESPACE file.
R CMD check --as-cran now also checks DOIs in package CITATION and Rd files.
R CMD Rdconv and R CMD Rd2pdf each have a new option --RdMacros=pkglist
which allows Rd macros to be specified before processing.


The previously included versions of zlib, bzip2, xz and PCRE have been removed,
so suitable external (usually system) versions are required (see the R Installation and
Administration manual).
The unexported and undocumented Windows-only devices cairo_bmp(),
cairo_png() and cairo_tiff() have been removed. (These devices should
be used as e.g. bmp(type = "cairo").)
(Windows only) Function setInternet2() has no effect and will be removed in due
course. The choice between methods "internal" and "wininet" is now made by the
method arguments of url() and download.file() and their defaults can be set via
options. The out-of-the-box default remains "wininet" (as it has been since R 3.2.2).
[<- with an S4 value into a list currently embeds the S4 object into its own list such
that the end result is roughly equivalent to using [[<-. That behavior is deprecated.
In the future, the S4 value will be coerced to a list with as.list().
Package tools functions package.dependencies(), pkgDepends(), etc are depre-
cated now, mostly in favor of package_dependencies() which is both more flexible
and efficient.


Support for very old versions of valgrind (e.g., 3.3.0) has been removed.
The included libtool script (generated by configure) has been updated to version
2.4.6 (from 2.2.6a).
libcurl version 7.28.0 or later with support for the https protocol is required for
installation (except on Windows).

BSD networking is now required (except on Windows) and so

capabilities("http/ftp") is always true.
configure uses pkg-config for PNG, TIFF and JPEG where this is available. This
should work better with multiple installs and with those using static libraries.
The minimum supported version of OS X is 10.6 (Snow Leopard): even that has
been unsupported by Apple since 2012.
The configure default on OS X is --disable-R-framework: enable this if you
intend to install under /Library/Frameworks and use with
The minimum preferred version of PCRE has since R 3.0.0 been 8.32 (released in Nov
2012). Versions 8.10 to 8.31 are now deprecated (with warnings from configure),
but will still be accepted until R 3.4.0.
configure looks for C functions __cospi, __sinpi and __tanpi and uses these if
cospi etc are not found. (OS X is the main instance.)
(Windows) R is now built using gcc 4.9.3. This build will require recompilation of at
least those packages that include C++ code, and possibly others. A build of R-devel
using the older toolchain will be temporarily available for comparison purposes.
During the transition, the environment variable R_COMPILED_BY has been defined to
indicate which toolchain was used to compile R (and hence, which should be used
to compile code in packages). The COMPILED_BY variable described below will be a
permanent replacement for this.
(Windows) A make and R CMD config variable named COMPILED_BY has been added.
This indicates which toolchain was used to compile R (and hence, which should be
used to compile code in packages).

The make macro AWK which used to be made available to files such as src/Makefile
is no longer set.

The API call logspace_sum introduced in R 3.2.0 is now remapped as an entry
point to Rf_logspace_sum, and its first argument has gained a const qualifier.
Code using it will need to be reinstalled.
Similarly, entry point log1pexp also defined in Rmath.h is remapped there to
R_GE_version has been increased to 11.
New API call R_orderVector1, a faster one-argument version of R_orderVector.
When R headers such as R.h and Rmath.h are called from C++ code in packages
they include the C++ versions of system headers such as <cmath> rather than the
legacy headers such as <math.h>. (Headers Rinternals.h and Rinterface.h
already did, and inclusion of system headers can still be circumvented by defining
NO_C_HEADERS, including as from this version for those two headers.)
The manual has long said that R headers should not be included within an
extern "C" block, and almost all the packages affected by this change were doing
Including header S.h from C++ code would fail on some platforms, and so gives a
compilation error on all.

The deprecated header Rdefines.h is now compatible with defining R_NO_REMAP.

The connections interface now includes a function R_GetConnection() which allows
packages implementing connections to convert R connection objects to Rconnection
handles. Code which previously used the low-level R-internal getConnection() entry
point should switch.

C-level asChar(x) is fixed for when x is not a vector, and it returns "TRUE"/"FALSE"
instead of "T"/"F" for logical vectors.
The first arguments of .colSums() etc (with an initial dot) are now named x rather
than X (matching colSums()): thus error messages are corrected.
A coef() method for class "maov" has been added to allow vcov() to work with
multivariate results. (PR#16380)
method = "libcurl" connections signal errors rather than retrieving HTTP error
pages (where the ISP reports the error). was not checking for unique row names; in particular, this
affected assignment to non-existing rows via numerical indexing. (PR#16570)
tail.matrix() did not work for zero rows matrices, and could produce row labels
such as "[1e+05,]".
Data frames with a column named "stringsAsFactors" now format and print cor-
rectly. (PR#16580)
cor() is now guaranteed to return a value with absolute value less than or equal to
1. (PR#16638)
Array subsetting now keeps names(dim(.)).
Blocking socket connection selection recovers more gracefully on signal interrupts.
The data.frame method of rbind() construction row.names works better in border-
line integer cases, but may change the names assigned. (PR#16666)
(X11 only) getGraphicsEvent() miscoded buttons and missed mouse motion events.
methods(round) now also lists round.POSIXt.
tar() now works with the default files = NULL. (PR#16716)
Jumps to outer contexts, for example in error recovery, now make intermediate
jumps to contexts where on.exit() actions are established instead of trying to run
all on.exit() actions before jumping to the final target. This unwinds the stack
gradually, releases resources held on the stack, and significantly reduces the chance
of a segfault when running out of C stack space. Error handlers established us-
ing withCallingHandlers() and options("error") specifications are ignored when
handling a C stack overflow error as attempting one of these would trigger a cascade
of C stack overflow errors. (These changes resolve PR#16753.)
The spacing could be wrong when printing a complex array. (Report and patch by
Lukas Stadler.)
pretty(d, n, min.n, *) for date-time objects d works again in border cases with
large min.n, returns a labels attribute also for small-range dates and in such cases
its returned length is closer to the desired n. (PR#16761) Additionally, it finally does
cover the range of d, as it always claimed.

tsp(x) <- NULL did not handle correctly objects inheriting from both "ts" and
"mts". (PR#16769)
install.packages() could give false errors when options("pkgType") was
"binary". (Reported by Jose Claudio Faria.)
A bug fix in R 3.0.2 fixed problems with locator() in X11, but introduced problems
in Windows. Now both should be fixed. (PR#15700)
download.file() with method = "wininet" incorrectly warned of download file
length difference when reported length was unknown. (PR#16805)
diag(NULL, 1) crashed because of missed type checking. (PR#16853)

format.POSIXlt() behaved incorrectly in R 3.2.4. E.g. the output of
format(as.POSIXlt(paste0(1940:2000,"-01-01"), tz = "CET"),usetz = TRUE)
ended in two "CEST" time formats.

install.packages() and related functions now give a more informative warning
when an attempt is made to install a base package.
summary(x) now prints with less rounding when x contains infinite values. (Request
of PR#16620.)
provideDimnames() gets an optional unique argument.
shQuote() gains type = "cmd2" for quoting in cmd.exe in Windows. (Response to
The data.frame method of rbind() gains an optional argument stringsAsFactors
(instead of only depending on getOption("stringsAsFactors")).
smooth(x, *) now also works for long vectors.
tools::texi2dvi() has a workaround for problems with the texi2dvi script supplied
by texinfo 6.1.
It extracts more error messages from the LaTeX logs when in emulation mode.

R CMD check will leave a log file build_vignettes.log from the re-building of
vignettes in the .Rcheck directory if there is a problem, and always if environment
variable _R_CHECK_ALWAYS_LOG_VIGNETTE_OUTPUT_ is set to a true value.


Use of SUPPORT_OPENMP from header Rconfig.h is deprecated in favour of the
standard OpenMP define _OPENMP.
(This has been the recommendation in the manual for a while now.)

The make macro AWK which is long unused by R itself but recorded in file
etc/Makeconf is deprecated and will be removed in R 3.3.0.
The C header file S.h is no longer documented: its use should be replaced by R.h.

kmeans(x, centers = <1-row>) now works. (PR#16623)
Vectorize() now checks for clashes in argument names. (PR#16577)
file.copy(overwrite = FALSE) would signal a successful copy when none had taken
place. (PR#16576)
ngettext() now uses the same default domain as gettext(). (PR#14605)
array(.., dimnames = *) now warns about non-list dimnames and, from R 3.3.0,
will signal the same error for invalid dimnames as matrix() has always done.
addmargins() now adds dimnames for the extended margins in all cases, as always
heatmap() evaluated its add.expr argument in the wrong environment. (PR#16583)
require() etc now give the correct entry of lib.loc in the warning about an old
version of a package masking a newer required one.
The internal deparser did not add parentheses when necessary, e.g. before [] or [[]].
(Reported by Lukas Stadler; additional fixes included as well).*, row.names=*) no longer produces corrupted data
frames from row names of incorrect length, but rather warns about them. This will
become an error.
url connections with method = "libcurl" are destroyed properly. (PR#16681)
withCallingHandler() now (again) handles warnings even during S4 generics argu-
ment evaluation. (PR#16111)
deparse(..., control = "quoteExpressions") incorrectly quoted empty expres-
sions. (PR#16686)
format()ting datetime objects ("POSIX[cl]?t") could segfault or recycle wrongly.
plot.ts(<matrix>, las = 1) now does use las.
saveRDS(*, compress = "gzip") now works as documented. (PR#16653)
(Windows only) The Rgui front end did not always initialize the console properly,
and could cause R to crash. (PR#16698)
dummy.coef.lm() now works in more cases, thanks to a proposal by Werner Sta-
hel (PR#16665). In addition, it now works for multivariate linear models ("mlm",
manova) thanks to a proposal by Daniel Wollschlaeger.
The as.hclust() method for "dendrogram"s failed often when there were ties in the
reorder() and midcache.dendrogram() now are non-recursive and hence applicable
to somewhat deeply nested dendrograms, thanks to a proposal by Suharto Anggono
in PR#16424.
cor.test() now calculates very small p values more accurately (affecting the result
only in extreme not statistically relevant cases). (PR#16704)
smooth(*, do.ends=TRUE) did not always work correctly in R versions between 3.0.0
and 3.2.3.

pretty(D) for date-time objects D now also works well if range(D) is (much) smaller
than a second. In the case of only one unique value in D, the pretty range now is more
symmetric around that value than previously.
Similarly, pretty(dt) no longer returns a length 5 vector with duplicated entries for
Date objects dt which span only a few days.
The figures in help pages such as ?points were accidentally damaged, and did not
appear in R 3.2.3. (PR#16708)
available.packages() sometimes deleted the wrong file when cleaning up temporary
files. (PR#16712)
The X11() device sometimes froze on Red Hat Enterprise Linux 6. It now waits
for MapNotify events instead of Expose events, thanks to Siteshwar Vashisht.
[dpqr]nbinom(*, size=Inf, mu=.) now works as limit case, for dpq as the Pois-
son. (PR#16727)
pnbinom() no longer loops infinitely in border cases.
approxfun(*, method="constant") and hence ecdf() which calls the former now
correctly predict NaN values as NaN. now displays NAs in Date columns in all cases. (PR#16709)

Some recently-added Windows time zone names have been added to the conversion
table used to convert these to Olson names. (Including those relating to changes for
Russia in Oct 2014, as in PR#16503.)
(Windows) Compatibility information has been added to the manifests for Rgui.exe,
Rterm.exe and Rscript.exe. This should allow win.version() and
to report the actual Windows version up to Windows 10.
Windows "wininet" FTP first tries EPSV / PASV mode rather than only using
active mode (reported by Dan Tenenbaum).
which.min(x) and which.max(x) may be much faster for logical and integer x and
now also work for long vectors.
The emulation part of tools::texi2dvi() has been somewhat enhanced, including
supporting quiet = TRUE. It can be selected by texi2dvi = "emulation".
(Windows) MiKTeX removed its texi2dvi.exe command in Sept 2015:
tools::texi2dvi() tries texify.exe if it is not found.
(Windows only) Shortcuts for printing and saving have been added to menus in
Rgui.exe. (Request of PR#16572.)
loess(..., iterTrace=TRUE) now provides diagnostics for robustness iterations,
and the print() method for summary(<loess>) shows slightly more.
The included version of PCRE has been updated to 8.38, a bug-fix release.
View() now displays nested data frames in a more friendly way. (Request with patch
in PR#15915.)



The included configuration code for libintl has been updated to that from gettext
version this should only affect how an external library is detected (and
the only known instance is under OpenBSD). (Wish of PR#16464.)
configure has a new argument --disable-java to disable the checks for Java.
The configure default for MAIN_LDFLAGS has been changed for the FreeBSD, NetBSD
and Hurd OSes to one more likely to work with compilers other than gcc (FreeBSD
10 defaults to clang).
configure now supports the OpenMP flags -fopenmp=libomp (clang) and
-qopenmp (Intel C).
Various macros can be set to override the default behaviour of configure when
detecting OpenMP: see file
Source installation on Windows has been modified to allow for MiKTeX installations
without texi2dvi.exe. See file MkRules.dist.

regexpr(pat, x, perl = TRUE) with Python-style named capture did not work cor-
rectly when x contained NA strings. (PR#16484)
The description of dataset ToothGrowth has been improved/corrected. (PR#15953)
model.tables(type = "means") and hence TukeyHSD() now support "aov" fits
without an intercept term. (PR#16437)
close() now reports the status of a pipe() connection opened with an explicit open
argument. (PR#16481)
Coercing a list without names to a data frame is faster if the elements are very long.
(Unix-only) Under some rare circumstances piping the output from Rscript or R -f
could result in attempting to close the input file twice, possibly crashing the process.
(Windows) was out of step with win.version() and did not report
Windows 8.
topenv(baseenv()) returns baseenv() again as in R 3.1.0 and earlier. This also fixes
compilerJIT(3) when used in .Rprofile.
detach()ing the methods package keeps .isMethodsDispatchOn() true, as long as
the methods namespace is not unloaded.
Removed some spurious warnings from configure about the preprocessor not finding
header files. (PR#15989)
rchisq(*, df=0, ncp=0) now returns 0 instead of NaN, and
dchisq(*, df=0, ncp=*) also no longer returns NaN in limit cases (where the
limit is unique). (PR#16521)
pchisq(*, df=0, ncp > 0, log.p=TRUE) no longer underflows (for ncp > 60).
nchar(x, "w") returned -1 for characters it did not know about (e.g. zero-width
spaces): it now assumes 1. It now knows about most zero-width characters and a few
more double-width characters.
Help for which.min() is now more precise about behavior with logical arguments.
The print width of character strings marked as "latin1" or "bytes" was in some
cases computed incorrectly.

abbreviate() did not give names to the return value if minlength was zero, unlike
when it was positive.
(Windows only) dir.create() did not always warn when it failed to create a direc-
tory. (PR#16537)
When operating in a non-UTF-8 multibyte locale (e.g. an East Asian locale on
Windows), grep() and related functions did not handle UTF-8 strings properly.
read.dcf() sometimes misread lines longer than 8191 characters. (Reported by Herve
Pages with a patch.)
within(df, ..) no longer drops columns whose name start with a ".".
The built-in HTTP server converted entire Content-Type to lowercase including pa-
rameters which can cause issues for multi-part form boundaries (PR#16541).
Modifying slots of S4 objects could fail when the methods package was not attached.
splineDesign(*, outer.ok=TRUE) (splines) is better now (PR#16549), and
interpSpline() now allows sparse=TRUE for speedup with non-small sizes.
If the expression in the traceback was too long, traceback() did not report the source
line number. (Patch by Kirill Muller.)
The browser did not truncate the display of the function when exiting with
options("deparse.max.lines") set. (PR#16581)
When bs(*, Boundary.knots=) had boundary knots inside the data range, extrap-
olation was somewhat off. (Patch by Trevor Hastie.)
var() and hence sd() warn about factor arguments which are deprecated now.
loess(*, weights = *) stored wrong weights and hence gave slightly wrong predic-
tions for newdata. (PR#16587)
aperm(a, *) now preserves names(dim(a)).
poly(x, ..) now works when either raw=TRUE or coef is specified. (PR#16597)
data(package=*) is more careful in determining the path.
prettyNum(*, decimal.mark, big.mark): fixed bug introduced when fixing

It is now easier to use secure downloads from https:// URLs on builds which
support them: no longer do non-default options need to be selected to do so. In
particular, packages can be installed from repositories which offer https:// URLs,
and those listed by setRepositories() now do so (for some of their mirrors).
Support for https:// URLs is available on Windows, and on other platforms if
support for libcurl was compiled in and if that supports the https protocol (system
installations can be expected to do). So https:// support can be expected except
on rather old OSes (an example being OS X Snow Leopard, where a non-system
version of libcurl can be used).

(Windows only) The default method for accessing URLs via download.file() and
url() has been changed to be "wininet" using Windows API calls. This changes the
way proxies need to be set and security settings made: there have been some reports
of ftp: sites being inaccessible under the new default method (but the previous
methods remain available).

cmdscale() gets new option list. for increased flexibility when a list should be
configure now supports texinfo version 6.0, which (unlike the change from 4.x to
5.0) is a minor update. (Wish of PR#16456.)
(Non-Windows only) download.file() with default method = "auto" now chooses
"libcurl" if that is available and a https:// or ftps:// URL is used.
(Windows only) setInternet2(TRUE) is now the default. The command-line option
--internet2 and environment variable R_WIN_INTERNET2 are now ignored.
Thus by default the "internal" method for download.file() and url() uses the
"wininet" method: to revert to the previous default use setInternet2(FALSE).
This means that https:// URLs can be read by default by download.file() (they
have been readable by file() and url() since R 3.2.0).
There are implications for how proxies need to be set (see ?download.file).
chooseCRANmirror() and chooseBioCmirror() now offer HTTPS mirrors in prefer-
ence to HTTP mirrors. This changes the interpretation of their ind arguments: see
their help pages.
capture.output() gets optional arguments type and split to pass to sink(), and
hence can be used to capture messages.

Header Rconfig.h now defines HAVE_ALLOCA_H if the platform has the alloca.h
header (it is needed to define alloca on Solaris and AIX, at least: see Writing R
Extensions for how to use it).


The libtool script generated by configure has been modified to support FreeBSD
>= 10 (PR#16410).

The HTML help page links to demo code failed due to a change in R 3.2.0.
If the na.action argument was used in model.frame(), the original data could be
modified. (PR#16436)
getGraphicsEvent() could cause a crash if a graphics window was closed while it
was in use. (PR#16438)
matrix(x, nr, nc, byrow = TRUE) failed if x was an object of type "expression".
strptime() could overflow the allocated storage on the C stack when the time-
zone had a non-standard format much longer than the standard formats. (Part of

options(OutDec = s) now signals a warning (which will become an error in the fu-
ture) when s is not a string with exactly one character, as that has been a documented
prettyNum() gains a new option input.d.mark which together with other changes,
e.g., the default for decimal.mark, fixes some format()ting variants with non-default
getOption("OutDec") such as in PR#16411.
download.packages() failed for type equal to either "both" or "binary". (Reported
by Dan Tenenbaum.)
The dendrogram method of labels() is much more efficient for large dendrograms,
now using rapply(). (Comment #15 of PR#15215)
The "port" algorithm of nls() could give spurious errors. (Reported by Radford
Reference classes that inherited from reference classes in another package could in-
validate methods of the inherited class. Fixing this requires adding the ability for
methods to be external, with the object supplied explicitly as the first argument,
named .self. See Inter-Package Superclasses in the documentation.
readBin() could fail on the SPARC architecture due to alignment issues. (Reported
by Radford Neal.)
qt(*, df=Inf, ncp=.) now uses the natural qnorm() limit instead of returning NaN.
Auto-printing of S3 and S4 values now searches for print() in the base namespace
and show() in the methods namespace instead of searching the global environment.
polym() gains a coefs = NULL argument and returns class "poly" just like poly()
which gets a new simple=FALSE option. They now lead to correct predict()ions,
e.g., on subsets of the original data.
rhyper(nn, <large>) now works correctly. (PR#16489)
ttkimage() did not (and could not) work so was removed. Ditto for tkimage.cget()
and tkimage.configure(). Added two Ttk widgets and missing subcommands
for Tks image command: ttkscale(), ttkspinbox(), tkimage.delete(),
tkimage.height(), tkimage.inuse(), tkimage.type(), tkimage.types(),
tkimage.width(). (PR#15372, PR#16450)
getClass("foo") now also returns a class definition when it is found in the cache
more than once.

utf8ToInt() now checks that its input is valid UTF-8 and returns NA if it is not.
install.packages() now allows type = "both" with repos = NULL if it can infer
the type of file.
nchar(x, *) and nzchar(x) gain a new argument keepNA which governs how the
result for NAs in x is determined. For nzchar() in general and nchar() in the R
3.2.x series, the default remains FALSE which is fully back compatible. From R 3.3.0,
nchar()s default will change to keepNA = NA and you are advised to consider this
for code portability.

news() more flexibly extracts dates from package NEWS.Rd files.

lengths(x) now also works (trivially) for atomic x and hence can be used more
generally as an efficient replacement of sapply(x, length) and similar.
The included version of PCRE has been updated to 8.37, a bug-fix release.
diag() no longer duplicates a matrix when extracting its diagonal.
as.character.srcref() gains an argument to allow characters corresponding to a
range of source references to be extracted.

acf() and ccf() now guarantee values strictly in [1, 1] (instead of sometimes very
slightly outside). PR#15832.
as.integer("111111111111") now gives NA (with a warning) as it does for the
corresponding numeric or negative number coercions. Further, as.integer(M + 0.1)
now gives M (instead of NA) when M is the maximal representable integer.
On some platforms nchar(x, "c") and nchar(x, "w") would return values (pos-
sibly NA) for inputs which were declared to be UTF-8 but were not, or for invalid
strings without a marked encoding in a multi-byte locale, rather than give an error.
Additional checks have been added to mitigate this.
apply(a, M, function(u) c(X = ., Y = .)) again has dimnames containing X
and Y (as in R < 3.2.0).
(Windows only) In some cases, the --clean option to R CMD INSTALL could fail.
(Windows only) choose.files() would occasionally include characters from the re-
sult of an earlier call in the result of a later one. (PR#16270)
A change in RSiteSearch() in R 3.2.0 caused it to submit invalid URLs. (PR#16329)
Rscript and command line R silently ignored incomplete statements at the end of a
script; now they are reported as parse errors. (PR#16350)
Parse data for very long strings was not stored. (PR#16354)
plotNode(), the workhorse of the plot method for "dendrogram"s is no longer recur-
sive, thanks to Suharto Anggono, and hence also works for deeply nested dendrograms.
The parser could overflow internally when given numbers in scientific format with
extremely large exponents. (PR#16358)
If the CRAN mirror was not set, install.packages(type = "both") and related
functions could repeatedly query the user for it. (Part of PR#16362)
The low-level functions .rowSums() etc. did not check the length of their argument,
so could segfault. (PR#16367)
The quietly argument of library() is now correctly propagated from
Under some circumstances using the internal PCRE when building R from source
would cause external libs such as -llzma to be omitted from the main link.
The .Primitive default methods of the logic operators, i.e., !, & and |, now give correct
error messages when appropriate, e.g., for `&`(TRUE) or `!`(). (PR#16385)
cummax(x) now correctly propagates NAs also when x is of type integer and begins
with an NA.

summaryRprof() could fail when the profile contained only two records. (PR#16395)
HTML vignettes opened using vignette() did not support links into the rest of the
HTML help system. (Links worked properly when the vignette was opened using
browseVignettes() or from within the help system.)
arima(*, xreg = .) (for d 1) computes estimated variances based on a the
number of effective observations as in R version 3.0.1 and earlier. (PR#16278)
slotNames(.) is now correct for "signature" objects (mostly used internally in
On some systems, the first string comparison after a locale change would result in NA.

anyNA() gains a recursive argument.
When x is missing and names is not false (including the default value),
Sys.getenv(x, names) returns an object of class "Dlist" and hence prints tidily.
(Windows.) shell() no longer consults the environment variable SHELL: too many
systems have been encountered where it was set incorrectly (usually to a path where
software was compiled, not where it was installed). R_SHELL, the preferred way to
select a non-default shell, can be used instead.
Some unusual arguments to embedFonts() can now be specified as character vectors,
and the defaults have been changed accordingly.
Functions in the Summary group duplicate less. (PR#15798)
(Unix-alikes.) system(cmd, input = ) now uses shell-execution-environment redi-
rection, which will be more natural if cmd is not a single command (but requires a
POSIX-compliant shell). (Wish of PR#15508)
read.fwf() and read.DIF() gain a fileEncoding argument, for convenience.
Graphics devices can add attributes to their description in .Device and .Devices.
Several of those included with R use a "filepath" attribute.
pmatch() uses hashing in more cases and so is faster at the expense of using more
memory. (PR#15697)
pairs() gains new arguments to select sets of variables to be plotted against each
other., extra_cols = FALSE) allows a minimal set of columns to be computed
on Unix-alikes: on some systems without properly-configured caching this can be
significantly faster with large file lists.
New function dir.exists() in package base to test efficiently whether one or more
paths exist and are directories.
dput() and friends gain new controls hexNumeric and digits17 which out-
put double and complex quantities as, respectively, binary fractions (exactly, see
sprintf("%a")) and as decimals with up to 17 significant digits.
save(), saveRDS() and serialize() now support ascii = NA which writes ASCII
files using sprintf("%a") for double/complex quantities. This is read-compatible
with ascii = TRUE but avoids binary->decimal->binary conversions with potential

loss of precision. Unfortunately the Windows C runtimes lack of C99 compliance

means that the format cannot be read correctly there in R before 3.1.2.
The default for formatC(decimal.mark =) has been changed to be
getOption("OutDec"); this makes it more consistent with format() and suit-
able for use in print methods, e.g. those for classes "density", "ecdf", "stepfun"
and "summary.lm".
getOption("OutDec") is now consulted by the print method for class "kmeans", by
cut(), dendrogram(), plot.ts() and quantile() when constructing labels and for
the report from legend(trace = TRUE).
(In part, wish of PR#15819.)
printNum() and hence format() and formatC() give a warning if big.mark and
decimal.mark are set to the same value (period and comma are not uncommonly
used for each, and this is a check that conventions have not got mixed).
merge() can create a result which uses long vectors on 64-bit platforms.
dget() gains a new argument keep.source which defaults to FALSE for speed (dput()
and dget() are most often used for data objects where this can make dget() many
times faster).
Packages may now use a file of common macro definitions in their help files, and may
import definitions from other packages.
A number of macros have been added in the new share/Rd directory for use in
package overview help pages, and promptPackage() now makes use of them.
tools::parse_Rd() gains a new permissive argument which converts unrecognized
macros into text. This is used by utils:::format.bibentry to allow LaTeX markup
to be ignored.
options(OutDec =) can now specify a multi-byte character, e.g.,
options(OutDec = "\u00b7") in a UTF-8 locale.
is.recursive(x) is no longer true when x is an external pointer, a weak reference
or byte code; the first enables all.equal(x, x) when x <- getClass(.).
ls() (aka objects()) and as.list.environment() gain a new argument sorted.
The "source" attribute (which has not been added to functions by R since before R
version 2.14.0) is no longer treated as special.
Function returnValue() has been added to give on.exit() code access to a func-
tions return value for debugging purposes.
crossprod(x, y) allows more matrix coercions when x or y are vectors, now
equalling t(x) %*% y in these cases (also reported by Radford Neal). Similarly,
tcrossprod(x,y) and %*% work in more cases with vector arguments.
Utility function dynGet() useful for detecting cycles, aka infinite recursions.
The byte-code compiler and interpreter include new instructions that allow many
scalar subsetting and assignment and scalar arithmetic operations to be handled
more efficiently. This can result in significant performance improvements in scalar
numerical code.
apply(m, 2, identity) is now the same as the matrix m when it has named row
A new function debuggingState() has been added, allowing to temporarily turn off

example() gets a new optional argument run.donttest and tools::Rd2ex() a corre-

sponding commentDonttest, with a default such that example(..) in help examples
will run \donttest code only if used interactively (a change in behaviour). gains an optional argument make.row.names, for potential
New function extSoftVersion() to report on the versions of third-party software in
use in this session. Currently reports versions of zlib, bzlib, the liblzma from xz,
PCRE, ICU, TRE and the iconv implementation.
A similar function grSoftVersion() in package grDevices reports on third-party
graphics software.
Function tcltk::tclVersion() reports the Tcl/Tk version.
Calling callGeneric() without arguments now works with primitive generics to some
vapply(x, FUN, FUN.VALUE) is more efficient notably for large length(FUN.VALUE);
as extension of PR#16061.
as.table() now allows tables with one or more dimensions of length 0 (such as
names(x) <- NULL now clears the names of call and ... objects.
library() will report a warning when an insufficient dependency version is masking
a sufficient one later on the library search path.
A new plot() method for class "raster" has been added.
New check_packages_in_dir_changes() function in package tools for conveniently
analyzing how changing sources impacts the check results of their reverse dependen-
Speed-up from Peter Haverty for ls() and methods:::.requirePackage() speeding
up package loading. (PR#16133)
New get0() function, combining exists() and get() in one call, for efficiency. gains an envir argument for specifying the environment from which
to retrieve the ... in the call, if any; this environment was wrong (or at least
undesirable) when the definition argument was a function.
topenv() has been made .Internal() for speedup, based on Peter Havertys proposal
in PR#16140.
getOption() no longer calls options() in the main case.
Optional use of libcurl (version 7.28.0 from Oct 2012 or later) for Internet access:
capabilities("libcurl") reports if this is available.
libcurlVersion() reports the version in use, and other details of the "libcurl"
build including which URL schemes it supports.
curlGetHeaders() retrieves the headers for http://, https://, ftp:// and
ftps:// URLs: analysis of these headers can provide insights into the existence
of a URL (it might for example be permanently redirected) and is so used in R CMD
check --as-cran.
download.file() has a new optional method "libcurl" which will handle more
URL schemes, follow redirections, and allows simultaneous downloads of multiple
url() has a new method "libcurl" which handles more URL schemes and follows
redirections. The default method is controlled by a new option url.method,

which applies also to the opening of URLs via file() (which happens implicitly
in functions such as read.table.)
When file() or url() is invoked with a https:// or ftps:// URL which the
current method cannot handle, it switches to a suitable method if one is available.
(Windows.) The DLLs internet.dll and internet2.dll have been merged. In
this version it is safe to switch (repeatedly) between the internal and Windows internet
functions within an R session.
The Windows internet functions are still selected by flag --internet2 or
setInternet2(). This can be overridden for an url() connection via its new method
download.file() has new method "wininet", selected as the default by
--internet2 or setInternet2().
parent.env<- can no longer modify the parent of a locked namespace or namespace
imports environment. Contributed by Karl Millar.
New function isNamespaceLoaded() for readability and speed.
names(env) now returns all the object names of an environment env, equivalently
to ls(env, all.names = TRUE, sorted = FALSE) and also to the names of the
corresponding list, names(as.list(env, all.names = TRUE)). Note that although
names() returns a character vector, the names have no particular ordering.
The memory manager now grows the heap more aggressively. This reduces the number
of garbage collections, in particular while data or code are loaded, at the expense of
slightly increasing the memory footprint.
New function trimws() for removing leading/trailing whitespace.
cbind() and rbind() now consider S4 inheritance during S3 dispatch and also obey
cbind() and rbind() will delegate recursively to methods::cbind2
(methods::rbind2) when at least one argument is an S4 object and S3 dispatch fails
(due to ambiguity).
(Windows.) download.file(quiet = FALSE) now uses text rather than Windows
progress bars in non-interactive use.
New function hsearch_db() in package utils for building and retrieving the help
search database used by, along with functions for inspecting the con-
cepts and keywords in the help search database.
New function .getNamespaceInfo(), a no-check version of getNamespaceInfo()
mostly for internal speedups.
The help search system now takes \keyword entries in Rd files which are not standard
keywords (as given in KEYWORDS in the R documentation directory) as concepts. For
standard keyword entries the corresponding descriptions are additionally taken as
New lengths() function for getting the lengths of all elements in a list.
New function toTitleCase() in package tools, tailored to package titles.
The matrix methods of cbind() and rbind() allow matrices as inputs which have
231 or more elements. (For cbind(), wish of PR#16198.)
The default method of image() has an explicit check for a numeric or logical matrix
(which was always required).

URLencode() will not by default encode further URLs which appear to be already
BIC(mod) and BIC(mod, mod2) now give non-NA numbers for arima() fitted models,
as nobs(mod) now gives the number of used observations for such models. This fixes
PR#16198, quite differently than proposed there.
The print() methods for "htest", "pairwise.htest" and "power.htest" objects
now have a digits argument defaulting to (a function of) getOption("digits"), and
influencing all printed numbers coherently. Unavoidably, this changes the display of
such test results in some cases.
Code completion for namespaces now recognizes all loaded namespaces, rather than
only the ones that are also attached.
The code completion mechanism can now be replaced by a user-specified completer
function, for (temporary) situations where the usual code completion is inappropriate.
unzip() will now warn if it is able to detect truncation when unpacking a file of 4GB
or more (related to PR#16243).
methods() reports S4 in addition to S3 methods; output is simplified when the class
argument is used. .S3methods() and methods::.S4methods() report S3 and S4
methods separately.
Higher order functions such as the apply functions and Reduce() now force arguments
to the functions they apply in order to eliminate undesirable interactions between lazy
evaluation and variable capture in closures. This resolves PR#16093.


The \donttest sections of Rs help files can be tested by
It is possible to request the use of system valgrind headers via configure op-
tion --with-system-valgrind-headers: note the possible future incompatibility
of such headers discussed in the R Installation and Administration manual. (Wish
of PR#16068.)
The included version of liblzma has been updated to xz-utils 5.0.7 (minor bug
fixes from 5.0.5).
configure options --with-system-zlib, --with-system-bzlib and
--with-system-pcre are now the default. For the time being there is fall-
back to the versions included in the R sources if no system versions are found or
(unlikely) if they are too old.
Linux users should check that the -devel or -dev versions of packages zlib,
bzip2/libbz2 and pcre as well as xz-devel/liblzma-dev (or similar names) are installed.
configure by default looks for the texi2any script from texinfo 5.1 or later, rather
than the makeinfo program. (makeinfo is a link to the Perl script texi2any in
texinfo 5.x.)
R CMD INSTALL gains an option --built-timestamp=STAMP allowing 100% repro-
ducible package building, thanks to Dirk Eddelbuettel.

There is support for testing the \dontrun and \donttest parts of examples in pack-

tools::testInstalledPackage() accepts new arguments

commentDontrun = FALSE and commentDonttest = FALSE.
R CMD check gains options --run-dontrun and --run-donttest.
The html generated by tools::Rd2HTML() and tools::toHTML() methods is now
xhtml 1.0 Strict.
The compiler packages utility function setCompilerOptions() now returns the old
values invisibly. The initial optimization level can also be set with the environment
R CMD build adds a NeedsCompilation field if one is not already present in the
R CMD check gains option --test-dir to specify an alternative set of tests to run.
R CMD check will now by default continue with testing after many types of errors,
and will output a summary count of errors at the end if any have occurred.
R CMD check now checks that the Title and Description fields are correctly
R CMD check --as-cran now:
checks a file can be processed: this needs pandoc installed.
checks the existence and accessibility of URLs in the DESCRIPTION, CITATION,
NEWS.Rd and files and in the help files (provided the build has
libcurl support).
reports non-ASCII characters in R source files when there is no package encoding
declared in the DESCRIPTION file.
reports (apparent) S3 methods exported but not registered.
reports overwriting registered S3 methods from base/recommended packages.
(Such methods are replaced in the affected package for the rest of the session,
even if the replacing namespace is unloaded.)
reports if the Title field does not appear to be in title case (see Writing R
Extensions: there may be false positives, but note that technical words should
be single-quoted and will then be accepted).
Most of these checks can also be selected by environment variables: see the R Inter-
nals manual.

New C API utility logspace_sum(logx[], n).
Entry points rbinom_mu, rnbinom_mu and rmultinom are remapped (by default) to
Rf_rbinom_mu etc. This requires packages using them to be re-installed.
.C(DUP = FALSE) and .Fortran(DUP = FALSE) are now ignored, so arguments are
duplicated if DUP = TRUE would do so. As their help has long said, .Call() is much
New entry point R_allocLD, like R_alloc but guaranteed to have sufficient alignment
for long double pointers.
isPairList() now returns TRUE for DOTSXP.

WINDOWS BUILD CHANGES: A number of changes to the Windows build system

are in development. The following are currently in place.

Installation using external binary distributions of zlib, bzip2, liblzma, pcre, libpng,
jpeglib and libtiff is now required, and the build instructions have been revised.
A new make target rsync-extsoft has been added to obtain copies of the external
libraries from cran.
Building the manuals now requires texi2any from texinfo 5.1 or later. cran binary
builds include the manuals, but by default builds from source will not, and they will
be accessed from cran. See the comments in src/gnuwin32/MkRules.dist for how
to specify the location of texi2any.
(Windows) Changes have been made to support an experimental Windows toolchain
based on GCC 4.9.2. The default toolchain continues to be based on GCC 4.6.3, as
the new toolchain is not yet stable enough. A change to a new toolchain is expected
during the R 3.2.x lifetime.

(Windows) The use of macro ZLIB_LIBS in file src/ (which has not
been documented for a long time) now requires an external libz.a to be available
(it is part of the goodies used to compile Windows binary packages). It would be
simpler to use -lz instead.
The default for option pkgType on platforms using binary packages is now "both",
so source packages will be tried if binary versions are not available or not up to date.
There are options for what install.packages(type = "both") (possibly called via
update.packages()) will do if compilation of a source package is desirable: see
?options (under utils).
If you intend not to accept updates as source packages, you should use
update.packages(type = "binary").


download.file(method = "lynx") is defunct.
Building R using the included versions of zlib, bzip2, xz and PCRE is deprecated:
these are frozen (bar essential bug-fixes) and will be removed for R 3.3.0.
The configure option --with-valgrind-instrumentation=3 has been withdrawn,
as it did not work with recent valgrind headers: it is now treated as level 2.
The MethodsList class in package methods had been deprecated in R 2.11.0 and is
defunct now. Functions using it are defunct if they had been deprecated in R 2.11.0,
and are deprecated now, otherwise.

Fixed two obscure bugs in pairlist subassignment, reported by Radford Neal as part
of pqR issue 16.
Fixes for bugs in handling empty arguments and argument matching by name in
all.equal() gains methods for environments and refClasses.
[<- and [[<- gain S4 data.frame methods to avoid corruption of S4 class information
by the S3 methods.
callNextMethod() should now work within a .local call when ... is absent from

dput(pairlist(x)) generates a call to the pairlist constructor instead of the list

Fix missing() when arguments are propagated through ... . (PR#15707)
eigen(m) now defaults to symmetric = TRUE even when the dimnames are asymmet-
ric if the matrix is otherwise symmetric. (PR#16151)
Fix issues with forwarding ... through callGeneric() and callNextMethod().
callGeneric() now works after a callNextMethod().
Subclass information is kept consistent when replacing an ordinary S4 class with
an old class via the S4Class argument to setOldClass(). Thus, for example, a
data.frame is valid for a list argument in the signature, and a factor is valid for
vector arguments.
In qbeta() the inversion of pbeta() is much more sophisticated. This works better
in corner cases some of which failed completely previously (PR#15755), or were using
too many iterations.
Auto-printing no longer duplicates objects when printing is dispatched to a method.
kmeans(x, k) would fail when nrow(x) >= 42949673. (Comment 6 of PR#15364)
Abbreviated locale-specific day and month names could have been truncated in those
rare locales where there are the same as the full names.
An irrelevant warning message from updating subclass information was silenced (the
namespace would not be writable in this case).

The internal method of download.file() can now handle files larger than 2GB on
32-bit builds which support such files (tested on 32-bit R running on 64-bit Windows).
kruskal.test() warns on more types of suspicious input.
The as.dendrogram() method for "hclust" objects gains a check argument pro-
tecting against memory explosion for invalid inputs.
capabilities() has a new item long.double which indicates if the build uses a
long double type which is longer than double.
nlm() no longer modifies the callback argument in place (a new vector is allocated for
each invocation, which mimics the implicit duplication that occurred in R < 3.1.0);
note that this is a change from the previously documented behavior. (PR#15958)
icuSetCollate() now accepts locale = "ASCII" which uses the basic C function
strcmp and so collates strings byte-by-byte in numerical order.
sessionInfo() tries to report the OS version in use (not just that compiled under,
and including details of Linux distributions).
model.frame() (used by lm() and many other modelling functions) now warns when
it drops contrasts from factors. (Wish of PR#16119)
install.packages() and friends now accept the value type = "binary" as a syn-
onym for the native binary type on the platform (if it has one).
Single source or binary files can be supplied for install.packages(type = "both")
and the appropriate type and repos = NULL will be inferred.

New function pcre_config() to report on some of the configuration options of the

version of PCRE in use. In particular, this reports if regular expressions using
\p{xx} are supported.
(Windows.) download.file(cacheOK = FALSE) is now supported when
internet2.dll is used.
browseURL() has been updated to work with Firefox 36.0 which has dropped support
for the -remote interface.


The included version of PCRE has been updated to 8.36.
configure accepts MAKEINFO=texi2any as another way to ensure texinfo 5.x is used
when both 5.x and 4.x are installed.

R CMD check now checks the packages used in \donttest sections of the examples
are specified in the DESCRIPTION file. (These are needed to run the examples inter-
R CMD check checks for the undeclared use of GNU extensions in Makefiles, and for
Makefiles with a missing final linefeed.
R CMD build will correct line endings in all Makefiles, not just those in the src
R CMD check notes uses of library() and require() in package code: see the section
Suggested packages of Writing R Extensions for good practice.


The configure option --with-valgrind-instrumentation=3 is deprecated and
will be removed in R 3.2.0.

(Windows.) Rscript.exe was missing a manifest specifying the modern style for
common controls (e.g., the download progress bar).
If a package had extra documentation files but no vignette, the HTML help system
produced an empty index page.
The parser now gives an error if a null character is included in a string using Unicode
escapes. (PR#16046)
qr.Q() failed on complex arguments due to pre-3.0(!) typo. (PR#16054)
abs() failed with named arguments when the argument was complex. (PR#16047)
"noquote" objects may now be used as columns in data frames. (PR#15997)
Some values with extremely long names were printed incorrectly. (PR#15999)
Extremely large exponents on zero expressed in scientific notation (e.g. 0.0e50000)
could give NaN. (PR#15976)
download.file() reported downloaded sizes as 0KB if less than 1MB, only for R
3.1.2 and only on big-endian platforms.
prompt() did not escape percent signs in the automatically generated usage section
of help files.

drop.terms() dropped some of the attributes of the object it was working with.
(Windows.) The command completion in Rgui.exe messed up the console.
(Windows.) The choose.files() command returned a blank string when the user
asked for a single file but cancelled the request. (PR#16074)
Math2 S4 group generics failed to correctly dispatch "structure"- and
"nonStructure"-derived classes.
loadNamespace() imposed undocumented restrictions on the versionCheck param-
eter. (Reported by Geoff Lee.)
Rare over-runs detected by AddressSanitizer in substr() and its replacement version
have been avoided.
Inter alia that fix gives the documented behaviour for substr(x, 1, 2) <- "" (sub-
sequently reported as PR#16214).
Loading packages incorrectly defining an S4 generic followed by a function of the same
name caused an erroneous cyclic namespace dependency error.
Declared vignette encodings are now always passed to the vignette engine.
Port Tomas Kaliberas fix from R-devel that restores the loadMethod() fast path,
effectively doubling the speed of S4 dispatch.
power.t.test() and power.prop.test() now make use of the extendInt option of
uniroot() and hence work in more extreme cases. (PR#15792)
If a package was updated and attached when its namespace was already loaded, it
could end up with parts from one version and parts from the other. (PR#16120)
tools:::.Rdconv() didnt accept --encoding= due to a typo. (PR#16121)
Unix-alike builds without a suitable makeinfo were documented to link the missing
HTML manuals to CRAN, but did not.
save(*, ascii=TRUE) and load() now correctly deal with NaNs. (PR#16137)
split.Date() retains fractional representations while avoiding incomplete class prop-
R_ext/Lapack.h had not been updated for changes made by LAPACK to the argu-
ment lists of its (largely internal) functions dlaed2 and dlaed3. (PR#16157)
RShowDoc("NEWS", "txt") had not been updated for the layout changes of R 3.1.0.
The xtfrm() method for class "Surv" has been corrected and its description ex-
mode(x) <- y would incorrectly evaluate x before changing its mode. (PR#16215)
besselJ(1, 2^64) and besselY(..) now signal a warning, returning NaN instead of
typically segfaulting. (Issue 3 of PR#15554)
HTML conversion of \href markup in .Rd files did not remove the backslash from
\% and so gave an invalid URL. In a related change, the \ escape is now required
in such URLs.


embedFonts() now defaults to format = "ps2write" for .ps and .eps files.
This is available in Ghostscript 9.x (since 2010) whereas the previous default,
format = "pswrite", was removed in Ghostscript 9.10.
For consistency with [dpqr]norm(), [dp]lnorm(sdlog = 0) model a point mass at
exp(mulog) rather than return NaN (for an error).
capabilities() now reports if ICU is compiled in for use for collation (it is only
actually used if a suitable locale is set for collation, and never for a C locale).
(OS X only.) Package tcltk checks when loaded if it is linked against the CRAN X11-
based Tcl/Tk and if so that the Tcl/Tk component and the X11 libraries are installed.
This allows more informative error messages to be given advising the installation of
the missing component or of XQuartz.
The X11() device and X11-based versions of the data editor and viewer (invoked by
edit() and View() for data frames and matrices from command-line R) check that
the X11 libraries are installed and if not advises installing XQuartz.
icuSetCollate() allows locale = "default", and locale = "none" to use OS ser-
vices rather than ICU for collation.
Environment variable R_ICU_LOCALE can be used to set the default ICU locale, in
case the one derived from the OS locale is inappropriate (this is currently necessary
on Windows).
New function icuGetCollate() to report on the ICU collation locale in use (if any).
utils::URLencode() was updated to use unreserved and reserved characters from
RFC 3986 ( instead of RFC 1738.
unique(warnings()) and c(warnings()) are now supported.
The Bioconductor version used by setRepositories() now defaults to 3.0. (It can
be set at runtime via environment variable R_BIOC_VERSION.)
Omegahat is no longer listed as providing Windows binary packages, e.g. by
setRepositories(). It has no binary packages available for R 3.1.x and those for
earlier versions were 32-bit only.


The configure script reports on the more important capabilities/options which will
not be compiled in.
More types of external BLAS are recognized by name in that report.
When building R as a shared library, the -L${R_HOME}/lib${R_ARCH} flag is placed
earlier in the link commands used during installation and when packages are installed:
this helps ensure that the current build has priority if an R shared library has already
been installed by e.g. install-libR in a library mentioned in LDFLAGS (and not in
your systems library directory as documented). (Wish of PR#15790.)
LaTeX package upquote is no longer required for Rs use of inconsolata.
(Windows only) If both 32- and 64-bit versions of R are installed, the bin/R.exe
and bin/Rscript.exe executables now run 64-bit R. (To run 32-bit R, overwrite
these files with copies of bin/i386/Rfe.exe.)

Running R CMD check with _R_CHECK_DEPENDS_ONLY_ true now makes the
VignetteBuilder packages available even if they are listed in Suggests, since
they are needed to recognise and process non-Sweave vignettes.

R CMD check now reports empty importFrom declarations in a NAMESPACE file, as

these are common errors (writing importFrom(Pkg) where import(Pkg) was in-
R CMD check now by default checks code usage directly on the package namespace
without loading and attaching the package and its suggests and enhances. For good
practice with packages in the Suggests field, see of Writing R Extensions.
For use of lazy-data objects in the packages own code, see ?data.

dmultinom() did not handle non-finite probabilities correctly.
prettyNum(x, zero.print=*) now also works when x contains NAs.
A longstanding bug exhibited by nlminb() on Windows was traced to a compiler bug
in gcc 4.6.3; a workaround has been put in place. (PR#15244 and PR#15914).
Rendering of \command in HTML versions of help pages has been improved: this is
particularly evident on the help page for INSTALL.
as.hexmode(x) and as.octmode(x) now behave correctly for some numeric x, e.g.,
c(NA, 1) or c(1, pi).
drop1() failed if the scope argument had no variables to drop. (PR#15935)
edit() (and hence fix()) failed if an object had a non-character attribute named
"source" (an attribute that had been used in R prior to version 2.14.0).
callGeneric() could fail if the generic had ... as a formal argument. (PR#15937).
Forking in package parallel called C entry point exit in the child. This was unsafe
(_exit should have been called), and could flush stdin of the main R process (seen
most often on Solaris).
As good practice, stdout is now flushed before forking a child.
R objects such as list(`a\b` = 1) now print correctly.
getAnywhere("C_pbinom") now returns correctly a single object (rather than unlist-
ing it).
The confint() method for nls() fits failed it these has specified parameter limits
despite using an algorithm other than "port". (PR#15960)
Subclassing an S4 class failed if the class required arguments to the generator, through
its initialize() method.
removeSource() did not properly handle expressions containing arguments that were
supplied as missing, e.g. x[i,]. (PR#15957)
as.environment(list()) now works, and as.list() of such an environment is now
the same as list().
Several tcltk functions failed when run in unusual environments. (PR#15970)
options(list()) now works (trivially). (PR#15979)
merge(<dendrogram>, ..) now works correctly for two independent dendrograms
(PR#15648), and still compatibly via adjust = "auto" e.g. for two branches of an
existing dendrogram.
The plot method for "hclust" objects gets an optional argument check; when that
is true (the default) it checks more carefully for valid input.
(Windows only) If a user chose to install 64 bit R but not 32 bit R, the bin/R and
bin/Rscript executables failed to run. (PR#15981)

Various possible buffer overruns have been prevented, and missed memory protection
added. (PR#15990)
Rscript no longer passes --args to R when there are no extra (user) arguments.
objects like getClass("refClass")@prototype now print() and str() without er-
identical() now also looks at the S4 bit.
hist(x, breaks) is more robust in adding a small fuzz to few breaks when some are
very large. (PR#15988)
sub() and gsub() did not handle regular expressions like "\s{2,}" properly if the
text contained NA or non-ASCII elements in a UTF-8 locale. Part of this was due to
a bug in the TRE library. (PR#16009)
RShowDoc("NEWS") now displays the PDF version.
Matrices and arrays with last dimension zero did not print at all or incompletely.
plot.histogram() and hence hist() now respect the xaxs, yaxs and lab graphics
parameters. (PR#16021)
bw.SJ(x) and other bw.*() no longer segfault when x contains non-finite values.
R CMD Rd2pdf unintentionally ignored its --os option.
The internal method of download.file() was not reporting file sizes and progress
correctly on files larger than 2GB (inherited from libxml2). This is corrected for
64-bit builds (32-bit platforms may not support such files, but where possible will be
supported in future versions of R).
Work around a bug in OS X Yosemite where key environment variables may be
duplicated causing issues in subprocesses. The duplicates are now removed on R
startup (via Rprofile). (PR#16042)
Adjust X11 auto-launch detection in DISPLAY on OS X to recognize latest XQuartz.

When attach() reports conflicts, it does so compatibly with library() by using
R CMD Sweave no longer cleans any files by default, compatibly with versions
of R prior to 3.1.0. There are new options --clean, --clean=default and
tools::buildVignette() and tools::buildVignettes() with clean = FALSE no
longer remove any created files. buildvignette() gains a keep argument for more
cleaning customization.
The Bioconductor version used by setRepositories() can now be set by environ-
ment variable R_BIOC_VERSION at runtime, not just when R is installed. (It has been
stated that Bioconductor will switch from version 2.14 to version 3.0 during the
lifetime of the R 3.1 series.)
Error messages from bugs in embedded Sexpr code in Sweave documents now report
the source location.

type.convert(), read.table() and similar read.*() functions get a new numerals

argument, specifying how numeric input is converted when its conversion to double
precision loses accuracy. The default value, "allow.loss" allows accuracy loss, as in
R versions before 3.1.0.
For some compilers, integer addition could overflow without a warning. Rs internal
code for both integer addition and subtraction is more robust now. (PR#15774)
The function determining the default number of knots for smooth.spline() is now
exported, as .nknots.smspl().
dbeta(, a,b), pbeta(), qbeta() and rbeta() are now defined also for a = 0, b = 0,
or infinite a and b (where they typically returned NaN before).
Many package authors report that the RStudio graphics device does not work correctly
with their packages use of The new option =
TRUE) replaces the RStudio override by the default device as selected by R itself, still
respecting environment variables R_INTERACTIVE_DEVICE and R_DEFAULT_DEVICE.
readRDS() now returns visibly.
Modifying internal logical scalar constants now results in an error instead of a warning.
install.packages(repos = NULL) now accepts http:// or ftp:// URLs of pack-
age archives as well as file paths, and will download as required. In most cases
repos = NULL can be deduced from the extension of the URL.
The warning when using partial matching with the $ operator on data frames is now
only given when options("warnPartialMatchDollar") is TRUE.
Package help requests like package?foo now try the package foo whether loaded or
General help requests now default to trying all loaded packages, not just those on the
search path.
Added a new function promptImport(), to generate a help page for a function that
was imported from another package (and presumably re-exported, or help would not
be needed).


configure option --with-internal-tzcode can now be used with variable
The included version of PCRE has been updated to 8.35.
There is a new target make uninstall-libR to remove an installed shared/static
make install-libR now works if a sub-architecture is used, although the user will
need to specify libdir differently for different sub-architectures.
There is more extensive advice on which LaTeX packages are required to install R or
to make package manuals (as done by R CMD check) in the Writing R Extensions
Compilers/linkers were handling the visibility control in src/extra/xz incon-
sistently (and apparently in some cases incorrectly), so it has been simplified.
(Windows) There is updated support for the use of ICU for collation: see the R
Installation and Administration Manual.


dbinom(x, n), pbinom(), dpois(), etc, are slightly less restrictive in checking if n is
integer-valued. (Wish of PR#15734.)
pchisq(x, df, ncp, log.p = TRUE) is more accurate and no longer underflows for
small x and ncp < 80, e.g, for pchisq(1e-5, df = 100, ncp = 1, log = TRUE).
(Based on PR#15635 and a suggestion by Roby Joehanes.)
The s (step into) command in the debugger would cause R to step into expressions
evaluated there, not just into functions being debugged. (PR#15770)
The C code used by strptime() rejected time-zone offsets of more than +1200 (+1245,
+1300 and +1400 can occur). (PR#15768)
(Windows only.) png(type = "cairo", antialias = "gray") was not accepted.
Use of save(..., envir=) with named objects could fail. (PR#15758)
Sweave() mis-parsed Sexpr expressions that contained backslashes. (PR#15779)
The return value from options(foo = NULL) was not the previous value of the option.
enc2utf8() and enc2native() did not always mark the encoding of the return values
when it was known.
dnbinom(x, size = <large>, mu, log = TRUE) no longer underflows to -Inf for
large mu, thanks to a suggestion from Alessandro Mammana (MPI MolGen, Berlin).
pbeta(x, a, b, log = TRUE) no longer behaves discontinuously
(in a small x-region) because of denormalized numbers. Also,
pbeta(1-1e-12, 1e30, 1.001, log=TRUE) now terminates in real time.
The "CRAN" filter (see available.packages()) no longer removes duplicates other
than of packages on CRAN, and does not fail if there is no CRAN repository in
The device listing from dev2bitmap() and bitmap() was truncated to 1000 charac-
ters: modern versions of GhostScript on most platforms have many more devices.
(Windows.) Commands such as Sys.which() and pipe() which needed to find the
full path to a command could segfault if the long path name was much longer than
the short path name (which Sys.which() returns), as the behaviour of the Windows
API call had changed.
R CMD build will fail with an error if one of the packages specified in the
VignetteBuilder field is not installed. (Without loading those packages it can-
not be ascertained which files are intended to be vignettes. This means that the
VignetteBuilder packages have to be installed for package checking too.) (Wish of
Misguided attempts to use chull() with non-finite points now give an error (related
to PR#15777).
For a formula with exactly 32 variables the 32nd variable was aliased to the inter-
cept in some C-level computations of terms, so that for example attempting to re-
move it would remove the intercept instead (and leave a corrupt internal structure).
anyDuplicated() silently returned wrong values when the first duplicate was at an
index which was too large to be stored in an integer vector (although a lot of RAM
and patience would have been needed to encounter this).
tools::Rd2ex(commentDontrun = FALSE) failed if the block had only one line.

Hexadecimal constants such as 0x110p-5L which were incorrectly qualified by L were

parsed incorrectly since R 3.0.0, with a slightly garbled warning. (PR#15753)
system() returned success on some platforms even if the system was unable to launch
a process. (PR#15796)
(Windows Rgui console.) Unbuffered output was sometimes not output immediately
if the prompt was not on the last line of the console.
The built-in help server did not declare the encoding for the DESCRIPTION or other
text files to be the package encoding, so non-ASCII characters could be displayed
R is now trying harder to not cleanup child processes that were not spawned by
mcparallel() on platforms that provide information about the source process of the
SIGCHLD signal. This allows 3rd party libraries to manage the exit status of children
that they spawn without R interfering.
mcmapply() was only parallelizing if the number of jobs was bigger than the number
of cores. It now parallelizes if the number of jobs is more than one.
Auto-printing would re-evaluate its argument when trying to dispatch to a print
method. This is now avoided when possible.
Unserializing (including load() and readRDS()) could silently return incorrect nu-
meric values from ASCII saves if there was a read error.
getParseData() could return incorrect values for the parents of some elements. (Re-
ported by Andrew Redd.)
Attempting to use data frames of 231 or more rows with merge() or to create a
merged data frame of that size now gives a clearer error message.
parse() did not check its file argument was a connection if it was not a character
string, so e.g. parse(FALSE) attempted to read from stdin.
Nor did dump() and dput().
The "help.try.all.packages" option was ignored when the shortcut syntax for
help was used, e.g. ?foo.
A potential segfault in string allocation has been fixed. (Found by Radford Neal.)
Potential memory protection errors in sort() and D() have been fixed. (Found by
Radford Neal.)
Fixed a lack of error checking in graphics event functions. (Found by Radford Neal;
a different patch used here than the one in pqR.)
numericDeriv() sometimes miscalculated the gradient. (PR#15849, reported origi-
nally by Radford Neal)

type.convert() (and hence by default read.table()) returns a character vector or
factor when representing a numeric input as a double would lose accuracy. Similarly
for complex inputs.
If a file contains numeric data with unrepresentable numbers of decimal places that
are intended to be read as numeric, specify colClasses in read.table() to be

tools::Rdiff(useDiff = FALSE) is closer to the POSIX definition of diff -b (as

distinct from the description in the man pages of most systems).
New function anyNA(), a version of any( which is fast for atomic vectors,
based on a proposal by Tim Hesterberg. (Wish of PR#15239.)
arrayInd(*, useNames = TRUE) and, analogously, which(*, arr.ind = TRUE)
now make use of names(.dimnames) when available.
is.unsorted() now also works for raw vectors.
The "table" method for (also useful as
now passes sep and base arguments to provideDimnames().
uniroot() gets new optional arguments, notably extendInt, allowing to auto-extend
the search interval when needed. The return value has an extra component,
switch(f, ...) now warns when f is a factor, as this typically happens accidentally
where the useR meant to pass a character string, but f is treated as integer (as always
The parser has been modified to use less memory.
The way the unary operators (+ - !) handle attributes is now more consistent. If
there is no coercion, all attributes (including class) are copied from the input to the
result: otherwise only names, dims and dimnames are.
colorRamp() and colorRampPalette() now allow non-opaque colours and a ramp in
opacity via the new argument alpha = TRUE. (Suggested by Alberto Krone-Martins,
but optionally as there are existing uses which expect only RGB values.) and get an optional vp.ex argument.
There is a new function find_gs_cmd() in the tools package to locate a GhostScript
executable. (This is an enhanced version of a previously internal function there.)
object.size() gains a format() method.
There is a new family, "ArialMT", for the pdf() and postscript() devices. This
will only be rendered correctly on viewers which have access to Monotype TrueType
fonts (which are sometimes requested by journals).
The text and PDF news files, including NEWS and NEWS.2, have been moved to the
doc directory.
combn(x, simplify = TRUE) now gives a factor result for factor input x (previously
user error). (Related to PR#15442.)
Added utils::fileSnapshot() and utils::changedFiles() functions to allow
snapshots and comparison of directories of files.
make.names(names, unique=TRUE) now tries to preserve existing names. (Sugges-
tion of PR#15452.)
New functions cospi(x), sinpi(x), and tanpi(x), for more accurate computation of
cos(pi*x), etc, both in R and the C API. Using these gains accuracy in some cases,
e.g., inside lgamma() or besselI(). (Suggested by Morten Welinder in PR#15529.)
print.table(x, zero.print = ".") now also has an effect when x is not integer-
There is more support to explore the systems idea of time-zone names.
Sys.timezone() tries to give the current system setting by name (and succeeds at
least on Linux, OS X, Solaris and Windows), and OlsonNames() lists the names in
the systems Olson database. Sys.timezone(location = FALSE) gives the previous

Platforms with a 64-bit time_t type are allowed to handle conversions between the
"POSIXct" and "POSIXlt" classes for date-times outside the 32-bit range (before
1902 or after 2037): the existing workarounds are used on other platforms. (Note
that time-zone information for post-2037 is speculative at best, and the OS services
are tested for known errors and so not used on OS X.)
Currently time_t is usually long and hence 64-bit on Unix-alike 64-bit platforms:
however in several cases the time-zone database is 32-bit. For R for Windows it is
64-bit (for both architectures as from this version).
The "save.defaults" option can include a value for compression_level. (Wish of
colSums() and friends now have support for arrays and data-frame columns with 231
or more elements.
as.factor() is faster when f is an unclassed integer vector (for example, when called
from tapply()).
fft() now works with longer inputs, from the 12 million previously supported up to
2 billion. (PR#15593)
Complex svd() now uses LAPACK subroutine ZGESDD, the complex analogue of the
routine used for the real case.
Sweave now outputs .tex files in UTF-8 if the input encoding is declared to be
UTF-8, regardless of the local encoding. The UTF-8 encoding may now be declared
using a LaTeX comment containing the string %\SweaveUTF8 on a line by itself.
file.copy() gains a argument.
Printing of date-times will make use of the time-zone abbreviation in use at the time,
if known. For example, for Paris pre-1940 this could be LMT, PMT, WET or WEST.
To enable this, the "POSIXlt" class has an optional component "zone" recording the
abbreviation for each element.
For platforms which support it, there is also a component "gmtoff" recording the
offset from GMT where known.
(On Windows, by default on OS X and optionally elsewhere.) The system C function
strftime has been replaced by a more comprehensive version with closer conformance
to the POSIX 2008 standard.
dnorm(x, log = FALSE) is more accurate (but somewhat slower) for |x| > 5; as
suggested in PR#15620.
Some versions of the tiff() device have further compression options.
read.table(), readLines() and scan() have a new argument to influence the treat-
ment of embedded nuls.
Avoid duplicating the right hand side values in complex assignments when possible.
This reduces copying of replacement values in expressions such as Z$a <- a0 and
ans[[i]] <- tmp: some package code has relied on there being copies.
Also, a number of other changes to reduce copying of objects; all contributed by or
based on suggestions by Michael Lawrence.
The fast argument of KalmanLike(), KalmanRun() and KalmanForecast() has been
replaced by update, which instead of updating mod in place, optionally returns the
updated model in an attribute "mod" of the return value.
arima() and makeARIMA() get a new optional argument SSinit, allowing the choice
of a different state space initialization which has been observed to be more reliable
close to non-stationarity: see PR#14682.

warning() has a new argument noBreaks., to simplify post-processing of output

with options(warn = 1).
pushBack() gains an argument encoding, to support reading of UTF-8 characters
using scan(), read.table() and related functions in a non-UTF-8 locale.
all.equal.list() gets a new argument use.names which by default labels differing
components by names (if they match) rather than by integer index. Saved R output
in packages may need to be updated.
The methods for all.equal() and attr.all.equal() now have argument
check.attributes after ... so it cannot be partially nor positionally matched (as
it has been, unintentionally).
A side effect is that some previously undetected errors of passing empty arguments
(no object between commas) to all.equal() are detected and reported.
There are explicit checks that check.attributes is logical, tolerance is numeric
and scale is NULL or numeric. This catches some unintended positional matching.
The message for all.equal.numeric() reports a "scaled difference" only for
scale != 1.
all.equal() now has a "POSIXt" method replacing the "POSIXct" method.
The "Date" and "POSIXt" methods of seq() allows by = "quarter" for complete-
ness (by = "3 months" always worked).
file.path() removes any trailing separator on Windows, where they are invalid
(although sometimes accepted). This is intended to enhance the portability of code
written by those using POSIX file systems (where a trailing / can be used to confine
path matching to directories).
New function agrepl() which like grepl() returns a logical vector.
fifo() is now supported on Windows. (PR#15600)
sort.list(method = "radix") now allows negative integers (wish of PR#15644).
Some functionality of print.ts() is now available in .preformat.ts() for more
mcparallel() gains an option detach = TRUE which allows execution of code inde-
pendently of the current session. It is based on a new estranged = TRUE argument
to mcfork() which forks child processes such that they become independent of the
parent process.
The pdf() device omits circles and text at extremely small sizes, since some viewers
were failing on such files.
The rightmost break for the "months", "quarters" and "years" cases of
hist.POSIXlt() has been increased by a day. (Inter alia, fixes PR#15717.)
The handling of DF[i,] <- a where i is of length 0 is improved. (Inter alia, fixes
hclust() gains a new method "ward.D2" which implements Wards method correctly.
The previous "ward" method is "ward.D" now, with the old name still working.
Thanks to research and proposals by Pierre Legendre.
The sunspot.month dataset has been amended and updated from the official source,
whereas the sunspots and sunspot.year datasets will remain immutable. The doc-
umentation and source links have been updated correspondingly.
The summary() method for "lm" fits warns if the fit is essentially perfect, as most of
the summary may be computed inaccurately (and with platform-dependent values).

Programmers who use summary() in order to extract just a component which will be
reliable (e.g., $cov.unscaled) should wrap their calls in suppressWarnings().


The included version of LAPACK has been updated to 3.5.0.
There is some support for parallel testing of an installation, by setting TEST_MC_CORES
to an integer greater than one to indicate the maximum number of cores to be used
in parallel. (It is worth specifying at least 8 cores if available.) Most of these require
a make program (such as GNU make and dmake) which supports the $MAKE -j nproc
Except on Windows: the tests of standard package examples in make check are done
in parallel. This also applies to running tools::testInstalledPackages().
The more time-consuming regression tests are done in parallel.
The package checks in make check-devel and make check-recommended are done
in parallel.
More of make check will work if recommended packages are not installed: but rec-
ommended packages remain needed for thorough checking of an R build.
The version of tzcode included in src/extra/tzone has been updated. (Formerly
used only on Windows.)
The included (64-bit) time-zone conversion code and Olson time-zone
database can be used instead of the system version: use configure option
--with-internal-tzcode. This is the default on Windows and OS X. (Note that
this does not currently work if a non-default rsharedir configure variable is used.)
(It might be necessary to set environment variable TZ on OSes where this is not
already set, although the system timezone is deduced correctly on at least Linux, OS
X and Windows.)
This option also switches to the version of strftime included in directory
configure now tests for a C++11-compliant compiler by testing some basic features.
This by default tries flags for the compiler specified by CXX, but an alternative
compiler, options and standard can be specified by variables CXX1X, CXX1XFLAGS
and CXX1XSTD (e.g., -std=gnu++11).
R can now optionally be compiled to use reference counting instead of the NAMED
mechanism by defining SWITCH_TO_REFCNT in Rinternals.h. This may become the
default in the future.
There is a new option --use-system-tre to use a suitable system tre library: at
present this means a version from their git repository, after corrections. (Wish of

The CRANextra repository is no longer a default repository on Windows: all the binary
versions of packages from CRAN are now on CRAN, although CRANextra contains
packages from Omegahat and elsewhere used by CRAN packages.
Only vignettes sources in directory vignettes are considered to be vignettes and
hence indexed as such.
In the DESCRIPTION file,
License: X11

is no longer recognized as valid. Use MIT or BSD_2_clause instead, both of which

need + file LICENSE.
For consistency, entries in .Rinstignore are now matched case-insensitively on all
Help for S4 methods with very long signatures now tries harder to split the description
in the Usage field to no more than 80 characters per line (some packages had over
120 characters).
R CMD INSTALL --build (not Windows) now defaults to the internal tar() unless
There is support for compiling C++11 code in packages on suitable platforms: see
Writing R Extensions.
Fake installs now install the contents of directory inst: some packages use this to
install e.g. C++ headers for use by other packages that are independent of the package
itself. Option --no-inst can be used to get the previous behaviour.

The behaviour of the code browser has been made more consistent, in part following
the suggestions in PR#14985.
Calls to browser() are now consistent with calls to the browser triggered by debug(),
in that Enter will default to n rather than c.
A new browser command s has been added, to step into function calls.
A new browser command f has been added, to finish the current loop or function.
Within the browser, the command help will display a short list of available commands.

Only vignettes sources in directory vignettes are considered to be vignettes by
R CMD check. That has been the preferred location since R 2.14.0 and is now oblig-
For consistency, R CMD build now matches entries in .Rbuildignore and
vignettes/.install_extras case-insensitively on all platforms (not just on Win-
checkFF() (called by R CMD check by default) can optionally check foreign func-
tion calls for consistency with the registered type and argument count. This is
the default for R CMD check --as-cran or can be enabled by setting environment
variable _R_CHECK_FF_CALLS_ to registration (but is in any case suppressed by
--install=no). Because this checks calls in which .NAME is an R object and not
just a literal character string, some other problems are detected for such calls.
Functions suppressForeignCheck() and dontCheck() have been added to allow
package authors to suppress false positive reports.
R CMD check --as-cran warns about a false value of the DESCRIPTION field
BuildVignettes for Open Source packages, and ignores it. (An Open Source pack-
age needs to have complete sources for its vignettes which should be usable on a
suitably well-equipped system).
R CMD check --no-rebuild-vignettes is defunct:
R CMD check --no-build-vignettes has been preferred since R 3.0.0.
R CMD build --no-vignettes is defunct:
R CMD build --no-build-vignettes has been preferred since R 3.0.0.

R CMD Sweave and R CMD Stangle now process both Sweave and non-Sweave vi-
gnettes. The tools::buildVignette() function has been added to do the same
tasks from within R.
The flags returned by R CMD config --ldflags and (where installed)
pkg-config --libs libR are now those needed to link a front-end against
the (shared or static) R library.
Sweave.sty has a new option [inconsolata].
R CMD check customizations such as _R_CHECK_DEPENDS_ONLY_ make available pack-
ages only in LinkingTo only for installation, and not for loading/runtime tests.
tools::checkFF() reports on .C and .Fortran calls with DUP = FALSE if argument
check_DUP is true. This is selected by R CMD check by default.
R CMD check --use-gct can be tuned to garbage-collect less frequently using
gctorture2() via the setting of environment variable _R_CHECK_GCT_N_.
Where supported, tools::texi2dvi() limits the number of passes tried to 20.

(Windows only) A function R_WaitEvent() has been added (with declaration in
headerR.h) to block execution until the next event is received by R.
Remapping in the Rmath.h header can be suppressed by defining
The remapping of rround() in header Rmath.h has been removed: use fround()
ftrunc() in header Rmath.h is now a wrapper for the C99 function trunc(), which
might as well be used in C code: ftrunc() is still needed for portable C++ code.
The never-documented remapping of prec() to fprec() in header Rmath.h has
been removed.
The included LAPACK subset now contains ZGESDD and ZGELSD.
The function LENGTH() now checks that it is only applied to vector arguments. How-
ever, in packages length() should be used. (In R itself LENGTH() is a macro without
the function overhead of length().)
Calls to SET_VECTOR_ELT() and SET_STRING_ELT() are now checked for indices which
are in-range: several packages were writing one element beyond the allocated length.
allocVector3 has been added which allows custom allocators to be used for individual
vector allocations.


chol(pivot = TRUE, LINPACK = TRUE) is defunct.
Arguments EISPACK for eigen() and LINPACK for chol(), chol2inv(), solve() and
svd() are ignored: LAPACK is always used.
.find.package() and .path.package() are defunct: only the versions without the
initial dot introduced in R 2.13.0 have ever been in the API.
Partial matching when using the $ operator on data frames now throws a warning and
may become defunct in the future. If partial matching is intended, replace foo$bar
by foo[["bar", exact = FALSE]].
The long-deprecated use of \synopsis in the Usage section of .Rd files has been
removed: such sections are now ignored (with a warning).

package.skeleton()s deprecated argument namespace has been removed.

Many methods are no longer exported by package stats. They are all registered on
their generic, which should be called rather than calling a method directly.
Functions readNEWS() and checkNEWS() in package tools are defunct.
download.file(method = "lynx") is deprecated.
.C(DUP = FALSE) and .Fortran(DUP = FALSE) are now deprecated, and may be
disabled in future versions of R. As their help has long said, .Call() is much preferred.
R CMD check notes such usages (by default).
The workaround of setting R_OSX_VALGRIND has been removed: it is not needed in
current valgrind.

Calling lm.wfit() with no non-zero weights gave an array-overrun in the Fortran
code and a not very sensible answer. It is now special-cased with a simpler answer
(no qr component).
Error messages involving non-syntactic names (e.g., as produced by `\r` when that
object does not exist) now encode the control characters. (Reported by Hadley Wick-
getGraphicsEvent() caused 100% usage of one CPU in Windows. (PR#15500)
nls() with no start argument may now work inside another function (scoping issue).
pbeta() and similar work better for very large (billions) ncp.
Where time zones have changed abbreviations over the years, the software tries to
more consistently use the abbreviation appropriate to the time or if that is un-
known, the current abbreviation. On some platforms where the C function localtime
changed the tzname variables the reported abbreviation could have been that of the
last time converted.
all.equal(list(1), identity) now works.
Bug fix for pushing viewports in grid (reported by JJ Allaire and Kevin Ushey).
NOTE for anyone poking around within the graphics engine display list (despite the
warnings not to) that this changes what is recorded by grid on the graphics engine
display list.
Extra checks have been added for unit resolution and conversion in grid, to catch
instances of division-by-zero. This may introduce error messages in existing code
and/or produce a different result in existing code (but only where a non-finite location
or dimension may now become zero).
Some bugs in TRE have been corrected by updating from the git repository. This
allows R to be installed on some platforms for which this was a blocker (PR#15087
suggests Linux on ARM and HP-UX).
? applied to a call to an S4 generic failed in several cases. (PR#15680)
The implicit S4 generics for primitives with ... in their argument list were incorrect.
Bug fixes to methods::callGeneric(). (PR#15691)
The bug fix to aggregrate() in PR#15004 introduced a new bug in the case of no
grouping variables. (PR#15699)
In rare cases printing deeply nested lists overran a buffer by one byte and on a few
platforms segfaulted. (PR#15679)

The dendrogram method of as.dendrogram() was hidden accidentally, (PR#15703),

and order.dendrogram(d) gave too much for a leaf d. (PR#15702)
R would try to kill processes on exit that have pids ever used by a child process
spawned by mcparallel even though the current process with that pid was not ac-
tually its child.
cophenetic() applied to a "dendrogram" object sometimes incorrectly returned a
"Labels" attribute with dimensions. (PR#15706)
printCoefmat() called from quite a few print() methods now obeys small
getOption("width") settings, line wrapping the "signif. codes" legend appro-
priately. (PR#15708)
model.matrix() assumed that the stored dimnames for a matrix was NULL or length
2, but length 1 occurred.
The clipping region for a device was sometimes used in base graphics before it was

On Windows there is support for making .texi manuals using texinfo 5.0 or later:
the setting is in file src/gnuwin32/MkRules.dist.
A packaging of the Perl script and modules for texinfo 5.2 has been made available
write.table() now handles matrices of 231 or more elements, for those with large
amounts of patience and disc space.
There is a new function, La_version(), to report the version of LAPACK in use.
The HTML version of An Introduction to R now has links to PNG versions of the
There is some support to produce manuals in ebook formats. (See
doc/manual/Makefile. Suggested by Mauro Cavalcanti.)
On a Unix-alike Sys.timezone() returns NA if the environment variable TZ is unset,
to distinguish it from an empty string which on some OSes means the UTC time zone.
The backtick may now be escaped in strings, to allow names containing them to be
constructed, e.g. `\``. (PR#15621)
read.table(), readLines() and scan() now warn when an embedded nul is found
in the input. (Related to PR#15625 which was puzzled by the behaviour in this
unsupported case.)
(Windows only.) file.symlink() works around the undocumented restriction of the
Windows system call to backslashes. (Wish of PR#15631.)
KalmanForecast(fast = FALSE) is now the default, and the help contains an ex-
ample of how fast = TRUE can be used in this version. (The usage will change in
strptime() now checks the locale only when locale-specific formats are used and
caches the locale in use: this can halve the time taken on OSes with slow system
functions (e.g., OS X).

strptime() and the format() methods for classes "POSIXct", "POSIXlt" and "Date"
recognize strings with marked encodings: this allows, for example, UTF-8 French
month names to be read on (French) Windows.
iconv(to = "utf8") is now accepted on all platforms (some implementations did
already, but GNU libiconv did not: however converted strings were not marked as
being in UTF-8). The official name, "UTF-8" is still preferred.
available.packages() is better protected against corrupt metadata files. (A recur-
ring problem with Debian package shogun-r: PR#14713.)
Finalizers are marked to be run at garbage collection, but run only at a somewhat
safer later time (when interrupts are checked). This circumvents some problems with
finalizers running arbitrary code during garbage collection (the known instances being
running options() and (C-level) path.expand() re-entrantly).


The included version of PCRE has been updated to 8.34. This fixes bugs and makes
the behaviour closer to Perl 5.18. In particular, the concept of space includes VT
and hence agrees with POSIXs.

The new field SysDataCompression in the DESCRIPTION file allows user control
over the compression used for sysdata.rda objects in the lazy-load database.
install.packages(dependencies = value) for value = NA (the default) or
value = TRUE omits packages only in LinkingTo for binary package installs.

The long undocumented remapping of rround() to Rf_fround() in header Rmath.h
is now formally deprecated: use fround() directly.
Remapping of prec() and trunc() in the Rmath.h header has been disabled in
C++ code (it has caused breakage with libc++ headers).

getParseData() truncated the imaginary part of complex number constants. (Re-
ported by Yihui Xie.)
dbeta(x, a, b) with a or b within a factor of 2 of the largest representable number
could infinite-loop. (Reported by Ioannis Kosmidis.)
provideDimnames() failed for arrays with a 0 dimension. (PR#15465)
rbind() and cbind() did not handle list objects correctly. (PR#15468)
replayPlot() now checks if it is replaying a plot from the same session.
rasterImage() and grid.raster() now give error on an empty (zero-length) raster.
(Reported by Ben North.)
plot.lm() would sometimes scramble the labels in plot type 5. (PR#15458 and
min() did not handle NA_character_ values properly. (Reported by Magnus Thor
(Windows only.) readRegistry() would duplicate default values for keys.

str(..., strict.width = "cut") did not handle it properly when more than one
line needed to be cut. (Reported by Gerrit Eichner.)
Removing subclass back-references when S4 classes were removed or their namespace
unloaded had several bugs (e.g., PR#15481).
aggregate() could fail when there were too many levels present in the by argument.
namespaceImportFrom() needed to detect primitive functions when checking for du-
plicated imports (reported by Karl Forner).
getGraphicsEvent() did not exit when a user closed the graphics window.
Errors in vignettes were not always captured and displayed properly. (PR#15495)
contour() could fail when dealing with extremely small z values. (PR#15454)
Several functions did not handle zero-length vectors properly, including browseEnv(),
format(), gl(), relist() and (E.g., PR#15499)
Sweave() did not restore the R output to the console if it was interrupted by a user
in the middle of evaluating a code chunk. (Reported by Michael Sumner.)
Fake installs of packages with vignettes work again.
Illegal characters in the input caused parse() (and thus source()) to segfault.
The nonsensical use of nmax = 1 in duplicated() or unique() is now silently ig-
qcauchy(p, *) is now fully accurate even when p is very close to 1. (PR#15521)
The validmu() and valideta() functions in the standard glm() families now also
report non-finite values, rather than failing.
Saved vignette results (in a file) were not being compared to the new
ones during R CMD check.
Double-clicking outside of the list box (e.g., on the scrollbar) of a Tk listbox widget
generated by tk_select.list() no longer causes the window to close. (PR#15407)
Improved handling of edge cases in parallel::splitindices(). (PR#15552)
HTML display of results from and ?? sometimes contained badly
constructed links.
c() and related functions such as unlist() converted raw vectors to invalid logical
vectors. (PR#15535)
(Windows only) When a call to system2() specified one of stdin, stdout or stderr
to be a file, but the command was not found (e.g., it contained its arguments, or the
program was not on the PATH), it left the file open and unusable until R terminated.
(Reported by Mathew McLean.)
The bmp() device was not recording res = NA correctly: it is now recorded as 72 ppi.
Several potential problems with compiler-specific behaviour have been identified using
the Undefined Behaviour Sanitizer in conjunction with the clang compiler.
hcl() now honours NA inputs (previously they were mapped to black).
Some translations in base packages were being looked up in the main catalog rather
than that for the package.
As a result of the 3.0.2 change about the last second before the epoch, most conver-
sions which should have given NA returned that time. (The platforms affected include
Linux and OS X, but not Windows nor Solaris.)

rowsum() has more support for matrices and data frames with 231 or more elements.
predict(<lm object>, interval = "confidence", scale = <something>) now
works. (PR#15564)
The bug fix in 3.0.2 for PR#15411 was too aggressive, and sometimes removed spaces
that should not have been removed. (PR#15583)
Running R code in a tcltk callback failed to set the busy flag, which will be needed
to tell OS X not to App Nap.
The code for date-times before 1902 assumed that the offset from GMT in 1902 was a
whole number of minutes: that was not true of Paris (as recorded on some platforms).
Using Sys.setlocale to set LC_NUMERIC to "C" (to restore the sane behavior) no
longer gives a warning.
deparse() now deparses complex vectors in a way that re-parses to the original values.
(PR#15534, patch based on code submitted by Alex Bertram.)
In some extreme cases (more than 1015 ) integer inputs to dpqrxxx() functions might
have been rounded up by one (with a warning about being non-integer). (PR#15624)
Plotting symbol pch = 14 had the triangle upside down on some devices (typically
screen devices). The triangle is supposed to be point up. (Reported by Bill Venables.)
getSrcref() did not work on method definitions if rematchDefinition() had been
KalmanForecast(fast = FALSE) reported a (harmless) stack imbalance.
The count of observations used by KalmanRun() did not take missing values into
In locales where the abbreviated name of one month is a partial match for the full
name of a later one, the %B format in strptime() could fail. An example was French
on OS X, where juin is abbreviated to jui and partially matches juillet. Simi-
larly for weekday names.
pbeta(x, a, b, log.p = TRUE) sometimes underflowed to zero for very small and
very differently sized a, b. (PR#15641)
approx() and approxfun() now handle infinite values with the "constant" method.
stripchart() again respects reversed limits in xlim and ylim. (PR#15664)

The NEWS files have been re-organized.
This file contains news for R >= 3.0.0: news for the 0.x.y, 1.x.y and 2.x.y re-
leases is in files NEWS.0, NEWS.1 and NEWS.2. The latter files are now installed
when R is installed. An HTML version of news from 2.10.0 to 2.15.3 is available as
sum() for integer arguments now uses an integer accumulator of at least 64 bits and
so will be more accurate in the very rare case that a cumulative sum exceeds 253
(necessarily summing more than 4 million elements).

The example() and tools::Rd2ex() functions now have parameters to allow them
to ignore \dontrun markup in examples. (Suggested by Peter Solymos.)
str(x) is considerably faster for very large lists, or factors with 100,000 levels, the
latter as in PR#15337.
col2rgb() now converts factors to character strings not integer codes (suggested by
Bryan Hanson).
tail(warnings()) now works, via the new `[` method.
There is now support for the LaTeX style file zi4.sty which has in some distribu-
tions replaced inconsolata.sty.
unlist(x) now typically returns all non-list xs unchanged, not just the vector ones.
Consequently, format(lst) now also works when the list lst has non-vector elements.
The tools::getVignetteInfo() function has been added to give information about
installed vignettes.
New assertCondition(), etc. utilities in tools, useful for testing.
Profiling now records non-inlined calls from byte-compiled code to BUILTIN functions.
Various functions in stats and elsewhere that use non-standard evaluation are now
more careful to follow the namespace scoping rules. E.g., stats::lm() can now find
stats::model.frame() even if stats is not on the search path or if some package
defines a function of that name.
If an invalid/corrupt .Random.seed object is encountered in the workspace it is ig-
nored with a warning rather than giving an error. (This allows R itself to rely on a
working RNG, e.g. to choose a random port.)
seq() and give more explicit error messages if called with invalid (e.g.,
NaN) inputs.
When parse() finds a syntax error, it now makes partial parse information available
up to the location of the error. (Request of Reijo Sund.)
Methods invoked by NextMethod() had a different dynamic parent to the generic.
This was causing trouble where S3 methods invoked via lazy evaluation could lose
track of their generic. (PR#15267)
Code for the negative binomial distribution now treats the case size == 0 as a one-
point distribution at zero.
abbreviate() handles without warning non-ASCII input strings which require no
read.dcf() no longer has a limit of 8191 bytes per line. (Wish of PR#15250.)
formatC(x) no longer copies the class of x to the result, to avoid misuse creating
invalid objects as in PR#15303. A warning is given if a class is discarded.
Dataset npk has been copied from MASS to allow more tests to be run without
recommended packages being installed.
The initialization of the regression coefficients for non-degenerate differenced mod-
els in arima() has been changed and in some examples avoids a local maximum.
termplot() now has an argument transform.x to control the display of individual
terms in the plot. (PR#15329)
format() now supports digits = 0, to display nsmall decimal places.
There is a new read-only par() parameter called "page", which returns a logical
value indicating whether the next call will start a new page.

Processing Sweave and Rd documents to PDF now renders backticks and single quotes
better in several instances, including in \code and \samp expressions.
utils::modifyList() gets a new argument keep.null allowing NULL components in
the replacement to be retained, instead of causing corresponding components to be
tools::pkgVignettes() gains argument check; if set to TRUE, it will warn when it
appears a vignette requests a non-existent vignette engine.

R CMD check --as-cran checks the line widths in usage and examples sections of the
package Rd files.
R CMD check --as-cran now implies --timings.
R CMD check looks for command gfile if a suitable file is not found. (Although
file is not from GNU, OpenCSW on Solaris installs it as gfile.)
R CMD build (with the internal tar) checks the permissions of configure and
cleanup files and adds execute permission to the recorded permissions for these
files if needed, with a warning. This is useful on OSes and file systems which do not
support execute permissions (notably, on Windows).
R CMD build now weaves and tangles all vignettes, so suggested packages are not
required during package installation if the source tarball was prepared with current
R CMD build.
checkFF() (used by R CMD check) does a better job of detecting calls from other
packages, including not reporting those where a function has been copied from another
namespace (e.g., as a default method). It now reports calls where .NAME is a symbol
registered in another package.
On Unix-alike systems, R CMD INSTALL now installs packages group writably when-
ever the library (lib.loc) is group writable. Hence, update.packages() works for
other group members (suggested originally and from a patch by Dirk Eddelbuettel).
R CMD javareconf now supports the use of symbolic links for JAVA_HOME on platforms
which have realpath. So it is now possible to use
R CMD javareconf JAVA_HOME=/usr/lib/jvm/java-1.7.0
on a Linux system and record that value rather than the frequently-changing full path
such as /usr/lib/jvm/java-1.7.0-openjdk-
(Windows only.) Rscript -e requires a non-empty argument for consistency with
Unix versions of R. (Also Rterm -e and R -e.)
R CMD check does more thorough checking of declared packages and namespaces. It
packages declared in more than one of the Depends, Imports, Suggests and
Enhances fields of the DESCRIPTION file.
namespaces declared in Imports but not imported from, neither in the
NAMESPACE file nor using the :: nor ::: operators.
packages which are used in library() or requires() calls in the R code but were
already put on the search path via Depends.
packages declared in Depends not imported via the NAMESPACE file (except the
standard packages). Objects used from Depends packages should be imported
to avoid conflicts and to allow correct operation when the namespace is loaded
but not attached.

objects imported via ::: calls where :: would do.

objects imported by :: which are not exported.
objects imported by ::: calls which do not exist.
See Writing R Extensions for good practice.
R CMD check optionally checks for non-standard top-level files and directories (which
are often mistakes): this is enabled for --as-cran.
LaTeX style file upquote.sty is no longer included (the version was several years
old): it is no longer used in R. A much later version is commonly included in LaTeX
distributions but does not play well with the ae fonts which are the default for Sweave
R CMD build makes more use of the build sub-directory of package sources, for
example to record information about the vignettes.
R CMD check analyses ::: calls.


The macros used for the texinfo manuals have been changed to work better with the
incompatible changes made in texinfo 5.x.
The minimum version for a system xz library is now 5.0.3 (was 4.999). This is in part
to avoid 5.0.2, which can compress in ways other versions cannot decompress.
The included version of PCRE has been updated to 8.33.
The included version of zlib has been updated to 1.2.8, a bug-fix release.
The included version of xz utilss liblzma has been updated to 5.0.5.
Since javareconf (see above) is used when R is installed, a stable link for JAVA_HOME
can be supplied then.
Configuring with --disable-byte-compilation will override the DESCRIPTION
files of recommended packages, which typically require byte-compilation.
More of the installation and checking process will work even when TMPDIR is set to a
path containing spaces, but this is not recommended and external software (such as
texi2dvi) may fail.

Installation is aborted immediately if a LinkingTo package is not installed.
R CMD INSTALL has a new option --no-byte-compile which will override a
ByteCompile field in the packages DESCRIPTION file.
License BSD is deprecated: use BSD_3_clause or BSD_2_clause instead.
License X11 is deprecated: use MIT or BSD_2_clause instead.
Version requirements for LinkingTo packages are now recognized: they are checked
at installation. (Fields with version requirements were previously silently ignored.)
The limit of 500 S3method entries in a NAMESPACE file has been removed.
The default version of Bioconductor for its packages has been changed to the up-
coming 2.13, but this can be set by the environment variable R_BIOC_VERSION when
R is installed.


Rdefines.h has been tweaked so it can be included in C++ code after

R_ext/Boolean.h (which is included by R.h).
Note that Rdefines.h is not kept up-to-date, and Rinternals.h is preferred for
new code.
eval and applyClosure are now protected against package code supplying an invalid


The unused namespace argument to package.skeleton() is now formally deprecated
and will be removed in R 3.1.0.
plclust() is deprecated: use the plot() method for class "hclust" instead.
Functions readNEWS() and checkNEWS() in package tools are deprecated (and they
have not worked with current NEWS files for a long time).

An Introduction to R has a new chapter on using R as a scripting language including
interacting with the OS.

help.request() could not determine the current version of R on CRAN. (PR#15241)
On Windows, failed on root directories unless the path was terminated
with an explicit ".". (PR#15302)
The regmatches<-() replacement function mishandled results coming from
regexpr(). (PR#15311)
The help for setClass() and representation() still suggested the deprecated ar-
gument representation=. (PR#15312)
R CMD config failed in an installed build of R 3.0.1 (only) when a sub-architecture
was used. (Reported by Berwin Turlach.)
On Windows, the installer modified the etc/Rconsole and etc/
files even when default options were chosen, so the MD5 sums did not refer to the
installed versions. (Reported by Tal Galili.)
plot(hclust(), cex =) respects cex again (and possibly others similarly). (Re-
ported by Peter Langfelder.)
If multiple packages were checked by R CMD check, and one was written for a different
OS, it would set --no-install for all following packages as well as itself.
qr.coef() and related functions did not properly coerce real vectors to complex when
necessary. (PR#15332)
ftable(a) now fixes up empty dimnames such that the result is printable.
package.skeleton() was not starting its search for function objects in the correct
place if environment was supplied. (Reported by Karl Forner.)
Parsing code was changing the length field of vectors and confusing the memory
manager. (PR#15345)
The Fortran routine ZHER2K in the reference BLAS had a comment-out bug in two
places. This caused trouble with eigen() for Hermitian matrices. (PR#15345 and
report from Robin Hankin)
vignette() and browseVignettes() did not display non-Sweave vignettes properly.

Two warning/error messages have been corrected: the (optional) warning produced by
a partial name match with a pairlist, the error message from a zero-length argument
to the : operator. (Found by Radford Neal; PR#15358, PR#15356)
svd() returned NULL rather than omitting components as documented. (Found by
Radford Neal; PR#15360)
mclapply() and mcparallel() with silent = TRUE could break a process that uses
stdout output unguarded against broken pipes (e.g., zip will fail silently). To work
around such issues, they now replace stdout with a descriptor pointed to /dev/null
instead. For this purpose, internal closeStdout and closeStderr functions have
gained the to.null flag.
log(), signif() and round() now raise an error if a single named argument is not
named x. (PR#15361)
deparse() now deparses raw vectors in a form that is syntactically correct.
The jpeg driver in Sweave created a JPEG file, but gave it a .png extension.
Deparsing of infix operators with named arguments is improved. (PR#15350)
mget(), and numericDeriv() did not duplicate arguments properly.
(PR#15352, PR#15353, PR#15354)
kmeans(algorithm = "Hartigan-Wong") now always stops iterating in the QTran
stage. (PR#15364).
read.dcf() re-allocated incorrectly and so could segfault when called on a file with
lines of more than 100 bytes.
On systems where mktime() does not set errno, the last second before the epoch
could not be converted from POSIXlt to POSIXct. (Reported by Bill Dunlap.)
add1.glm() miscalculated F-statistics when df > 1. (Bill Dunlap, PR#15386).
stem() now discards infinite inputs rather than hanging. (PR#15376)
The parser now enforces C99 syntax for floating point hexadecimal constants
(e.g., 0x1.1p0), rather than returning unintended values for malformed constants.
model.matrix() now works with very long LHS names (more than 500 bytes).
integrate() reverts to the pre-2.12.0 behaviour: from 2.12.0 to 3.0.1 it sometimes
failed to achieve the requested tolerance and reported error estimates that were ex-
ceeded. (PR#15219)
strptime() now handles %W fields with value 0. (PR#15915)
R is now better protected against people trying to interact with the console in startup
code. (PR#15325)
Subsetting 1D arrays often lost dimnames (PR#15301).
Unary + on a logical vector did not coerce to integer, although unary - did.
na.omit() and na.exclude() added a row to a zero-row data frame. (PR#15399)
All the (where necessary cut-down) vignettes are installed if R was configured with
source() did not display filenames when reporting syntax errors.
Syntax error reports misplaced the caret pointing out the bad token.

(Windows only) Starting R with R (instead of Rterm or Rgui) would lose any zero-
length strings from the command line arguments. (PR#15406)
Errors in the encoding specified on the command line via --encoding=foo were not
handled properly. (PR#15405)
If x is a symbol, is.vector(x, "name") now returns TRUE, since "name" and
"symbol" should be synonyms. (Reported by Herve Pages.)
R CMD rtags works on platforms (such as OS X) with a XSI-conformant shell com-
mand echo. (PR#15231)
is.unsorted(NA) returns false as documented (rather than NA).
R CMD LINK did not know about sub-architectures.
system() and system2() are better protected against users who misguidedly have
spaces in the temporary directory path. and edit() are now more likely to work on file paths containing spaces.
(Where external utilities are used, not the norm on Windows nor in which
should previously have worked.)
Packages using the methods package are more likely to work when they import it but
it is not attached. (Several parts of its C code were looking for its R functions on the
search path rather than in its namespace.)
lgamma(-x) is no longer NaN for very small x.
(Windows) system2() now respects specifying stdout and stderr as files if called
from Rgui. (PR#15393)
Closing an x11() device whilst locator() or identify() is in progress no longer
hangs R. (PR#15253)
list.dirs(full.names = FALSE) was not implemented. (PR#15170)
format() sometimes added unnecessary spaces. (PR#15411)
all.equal(check.names = FALSE) would ignore the request to ignore the names
and would check them as attributes.
The symbol set by tools::Rd2txt_options(itemBullet=) was not respected in
some locales. (PR#15435)
mcMap() was not exported by package parallel. (PR#15439)
plot() for TukeyHSD objects did not balance dev.hold() and dev.flush() calls on
multi-page plots. (PR#15449)

chooseCRANmirror() and chooseBioCmirror() gain an ind argument (like
mcparallel has a new argument mc.interactive which can modify the interactive
flag in the child process. The new default is FALSE which makes child processes non-
interactive by default (this prevents lock-ups due to children waiting for interactive
scan() now warns when end-of-file occurs within a quoted string.

count.fields() is now consistent with scan() in its handling of newlines in quoted

strings. Instead of triggering an error, this results in the current line receiving NA as
the field count, with the next line getting the total count of the two lines.
The default method of image() will plot axes of the class of xlim and ylim (and
hence of x and y if there is a suitable range() method). Based on a suggestion of
Michael Sumner.
load() now has a verbose argument for debugging support, to print the names of
objects just before loading them.
When loading a serialized object encounters a reference to a namespace which cannot
be loaded, this is replaced by a reference to the global environment, with a warning.
pairs() gains a line.main option for title placement.
The remaining instances in which serialization to a raw vector was limited to 2GB
have been unlimited on a 64-bit platform, and in most cases serialization to a vector
of more than 1GB will be substantially faster.

R CMD config now make use of personal Makevars files under ~/.R and a site file, in the same way as R CMD SHLIB and R CMD INSTALL. This makes
the utility more useful in package configure scripts.
On Windows finding the personal files may require the environment variable HOME set.
The old behaviour can be obtained with the new options --no-user-files and

Alternatives to the site and user customization files and
~/.R/Makevars can be specified via the environment variables R_MAKEVARS_SITE
and R_MAKEVARS_USER respectively. These can be used to suppress the use of the
default files by setting an empty value (where possible) or a non-existent path.

sys.source() did not report error locations when keep.source = TRUE.
as.POSIXct.numeric was coercing origin using the tz argument and not "GMT" as
documented (PR#14973).
The active binding to assign fields in reference classes has been cleaned up to reduce
dependence on the class package environment, also fixing bug in initializing read-only
fields (inspired by a report from Hadley Wickham).
str(d) no longer gives an error when names(d) contain illegal multibyte strings
Profiling of built-in functions with line.profiling= TRUE did not record the
line from which they were called.
citation(pkg) dropped the header and footer specified in the CITATION file
Quotes were handled differently when reading the first line and reading the rest, so
read.table() misread some files that contained quote characters (PR#15245).
cat() with sep a character vector of length greater than one and more than one
argument was using separators inconsistently (PR#15261).

On Windows in R 3.0.0, savePlot() failed because of an incorrect check on the

argument count.
unzip(list = TRUE) returned Names as a factor and not a character vector (as doc-
umented) for the internal method. (Noticed by Sean ORiordain.)
contourLines() now checks more comprehensively for conformance of its x, y and z
arguments (it was used incorrectly in package R2G2).
Saved graphics display lists are R version-specific. Attempting to load workspaces
containing them (or some other version-specific objects) aborted the load in R 3.0.0
and earlier; now it does a partial load and generates a warning instead.
In R 3.0.0, identify() and locator() did not record information correctly, so re-
playing a graph (e.g., by copying it to another device) would fail. (PR#15271)
Calling file.copy() or dirname() with the invalid input "" (which was being used
in packages, despite not being a file path) could have caused a segfault.
dirname("") is now "" rather than "." (unless it segfaulted).
supsmu() could read/write outside its input vectors for very short inputs (seen in
package rms for n = 4).
as.dendrogram()s hclust method uses less memory and hence gets considerably
faster for large (n 1000) clusterings, thanks to Daniel Mullner. (PR#15174)
The return value when all workers failed from
parallel::mclapply(mc.preschedule = TRUE) was a list of strings and not
of error objects. (Spotted by Karl Forner and Bernd Bischl.)
In R 3.0.0, when help() found multiple pages with the same alias, the HTML display
of all the selections was not produced. (PR#15282)
splinefun(method="monoH.FC") now produces a function with first argument named
x and allows deriv=3, as documented. (PR#15273)
summaryRprof() would only read the first chunksize lines of an Rprof file produced
with line.profiling=TRUE. By default, this is the first 100 seconds. (PR#15288)
lsfit() produced an incorrect error message when argument x had more columns
than rows or x had a different number of rows than y. (Spotted by Renaud Gaujoux.)
Binary operations on equal length vectors copied the class name from the second
operand when the first had no class name, but did not set the object bit. (PR#15299)
The trace() method for reference generator objects failed after those objects became
function definitions.
write.table() did not check that factors were constructed correctly, and so caused
a segment fault when writing bad ones. (PR#15300)
The internal HTTP server no longer chokes on POST requests without body. It will
also pass-through other request types for custom handlers (with the method stored
in Request-Method header) instead of failing.

Packages need to be (re-)installed under this version (3.0.0) of R.

There is a subtle change in behaviour for numeric index values 231 and larger. These
never used to be legitimate and so were treated as NA, sometimes with a warning. They
are now legal for long vectors so there is no longer a warning, and x[2^31] <- y will
now extend the vector on a 64-bit platform and give an error on a 32-bit one.
It is now possible for 64-bit builds to allocate amounts of memory limited only by
the OS. It may be wise to use OS facilities (e.g., ulimit in a bash shell, limit in
csh), to set limits on overall memory consumption of an R process, particularly in a
multi-user environment. A number of packages need a limit of at least 4GB of virtual
memory to load.
64-bit Windows builds of R are by default limited in memory usage to the amount of
RAM installed: this limit can be changed by command-line option --max-mem-size
or setting environment variable R_MAX_MEM_SIZE.
Negative numbers for colours are consistently an error: previously they were some-
times taken as transparent, sometimes mapped into the current palette and sometimes
an error.

identical() has a new argument, ignore.environment, used when comparing func-
tions (with default FALSE as before).
There is a new option, options(CBoundsCheck=), which controls how .C() and
.Fortran() pass arguments to compiled code. If true (which can be enabled by
setting the environment variable R_C_BOUNDS_CHECK to yes), raw, integer, double
and complex arguments are always copied, and checked for writing off either end of
the array on return from the compiled code (when a second copy is made). This also
checks individual elements of character vectors passed to .C().
This is not intended for routine use, but can be very helpful in finding segfaults in
package code.
In layout(), the limits on the grid size have been raised (again).
New simple provideDimnames() utility function.
Where methods for length() return a double value which is representable as an
integer (as often happens for package Matrix), this is converted to an integer.
Matrix indexing of data frames by two-column numeric indices is now supported for
replacement as well as extraction.
setNames() now has a default for its object argument, useful for a character result.
StructTS() has a revised additive constant in the loglik component of the result:
the previous definition is returned as the loglik0 component. However, the help
page has always warned of a lack of comparability of log-likelihoods for non-stationary
models. (Suggested by Jouni Helske.)
The logic in aggregate.formula() has been revised. It is now possible to use a
formula stored in a variable; previously, it had to be given explicitly in the function
install.packages() has a new argument quiet to reduce the amount of output
Setting an element of the graphics argument lwd to a negative or infinite value is now
an error. Lines corresponding to elements with values NA or NaN are silently omitted.
Previously the behaviour was device-dependent.

Setting graphical parameters cex, col, lty, lwd and pch in par() now requires a
length-one argument. Previously some silently took the first element of a longer
vector, but not always when documented to do so.
Sys.which() when used with inputs which would be unsafe in a shell (e.g., absolute
paths containing spaces) now uses appropriate quoting.
as.tclObj() has been extended to handle raw vectors. Previously, it only worked in
the other direction. (Contributed by Charlie Friedemann, PR#14939.)
New functions cite() and citeNatbib() have been added, to allow generation of
in-text citations from "bibentry" objects. A cite() function may be added to
bibstyle() environments.
A sort() method has been added for "bibentry" objects.
The bibstyle() function now defaults to setting the default bibliography style. The
getBibstyle() function has been added to report the name of the current default
scatter.smooth() now has an argument lpars to pass arguments to lines().
pairs() has a new log argument, to allow some or all variables to be plotted on
logarithmic scale. (In part, wish of PR#14919.)
split() gains a sep argument.
termplot() does a better job when given a model with interactions (and no longer
attempts to plot interaction terms).
The parser now incorporates code from Romain Francois parser package, to sup-
port more detailed computation on the code, such as syntax highlighting, comment-
based documentation, etc. Functions getParseData() and getParseText() access
the data.
There is a new function rep_len() analogous to for when speed is required
(and names are not).
The undocumented use rep(NULL, length.out = n) for n > 0 (which returns NULL)
now gives a warning.
demo() gains an encoding argument for those packages with non-ASCII demos: it
defaults to the package encoding where there is one.
strwrap() converts inputs with a marked encoding to the current locale: previously
it made some attempt to pass through as bytes inputs invalid in the current locale.
Specifying both rate and scale to [dpqr]gamma is a warning (if they are essentially
the same value) or an error.
merge() works in more cases where the data frames include matrices. (Wish of
optimize() and uniroot() no longer use a shared parameter object across calls.
(nlm(), nlminb() and optim() with numerical derivatives still do, as documented.)
The all.equal() method for date-times is now documented: times are regarded as
equal (by default) if they differ by up to 1 msec.
duplicated() and unique() gain a nmax argument which can be used to make them
much more efficient when it is known that there are only a small number of unique
entries. This is done automatically for factors.
Functions rbinom(), rgeom(), rhyper(), rpois(), rnbinom(), rsignrank() and
rwilcox() now return integer (not double) vectors. This halves the storage require-
ments for large simulations.

sort(), and sort.list() now use radix sorting for factors of less than
100,000 levels when method is not supplied. So does order() if called with a single
factor, unless na.last = NA.
diag() as used to generate a diagonal matrix has been re-written in C for speed and
less memory usage. It now forces the result to be numeric in the case diag(x) since
it is said to have zero off-diagonal entries.
backsolve() (and forwardsolve()) are now internal functions, for speed and sup-
port for large matrices.
More matrix algebra functions (e.g., chol() and solve()) accept logical matrices
(and coerce to numeric). has some support for n 231 : see its help for the limitations.
A different algorithm is used for (n, size, replace = FALSE,prob = NULL) for
n > 1e7 and size <= n/2. This is much faster and uses less memory, but does give
different results.
approxfun() and splinefun() now return a wrapper to an internal function in the
stats namespace rather than a .C() or .Call() call. This is more likely to work if
the function is saved and used in a different session.
The functions .C(), .Call(), .External() and .Fortran() now give an error (rather
than a warning) if called with a named first argument.
Sweave() by default now reports the locations in the source file(s) of each chunk.
clearPushBack() is now a documented interface to a long-existing internal call.
aspell() gains filters for R code, Debian Control Format and message catalog files,
and support for R level dictionaries. In addition, package utils now provides functions
aspell_package_R_files() and aspell_package_C_files() for spell checking R
and C level message strings in packages.
bibentry() gains some support for incomplete entries with a crossref field.
gray() and gray.colors() finally allow alpha to be specified.
monthplot() gains parameters to control the look of the reference lines. (Suggestion
of Ian McLeod.)
Added support for new %~% relation (is distributed as) in plotmath.
domain = NA is accepted by gettext() and ngettext(), analogously to stop() etc.
termplot() gains a new argument plot = FALSE which returns information to allow
the plots to be modified for use as part of other plots, but does not plot them.
(Contributed by Terry Therneau, PR#15076.), formerly an undocumented part of, is now available to copy a
device to a quartz() device. dev.copy2pdf() optionally does this for PDF output: defaults to PNG.
The default method of pairs() now allows text.panel = NULL and the use of
<foo>.panel = NULL is now documented.
setRefClass() and getRefClass() now return class generator functions, similar to
setClass(), but still with the reference fields and methods as before (suggestion of
Romain Francois).
New functions bitwNot(), bitwAnd(), bitwOr() and bitwXor(), using the internal
interfaces previously used for classes "octmode" and "hexmode".
Also bitwShiftL() and bitwShiftR() for shifting bits in elements of integer vectors.

New option "deparse.cutoff" to control the deparsing of language objects such as

calls and formulae when printing. (Suggested by a comment of Sarah Goslee.)
colors() gains an argument distinct.
New demo(colors) and demo(hclColors), with utility functions.
list.files() (aka dir()) gains a new optional argument no.. which allows to
exclude "." and ".." from listings.
Multiple time series are also of class "matrix"; consequently, head(), e.g., is more
encodeString() preserves UTF-8 marked encodings. Thus if factor levels are marked
as UTF-8 an attempt is made to print them in UTF-8 in RGui on Windows.
readLines() and scan() (and hence read.table()) in a UTF-8 locale now discard
a UTF-8 byte-order-mark (BOM). Such BOMs are allowed but not recommended by
the Unicode Standard: however Microsoft applications can produce them and so they
are sometimes found on websites.
The encoding name "UTF-8-BOM" for a connection will ensure that a UTF-8 BOM is
mapply(FUN, a1, ..) now also works when a1 (or a further such argument) needs a
length() method (which the documented arguments never do). (Requested by Herve
Pages; with a patch.)
.onDetach() is supported as an alternative to .Last.lib. Unlike .Last.lib, this
does not need to be exported from the packages namespace.
The srcfile argument to parse() may now be a character string, to be used in error
The format() method for ftable objects gains a method argument, propagated to
write.ftable() and print(), allowing more compact output, notably for LaTeX
formatting, thanks to Marius Hofert.
The function has been added to trigger immediate event
Sys.which() now returns NA (not "") for NA inputs (related to PR#15147).
The print() method for class "htest" gives fewer trailing spaces (wish of
Also print output from HoltWinters(), nls() and others.
loadNamespace() allows a version specification to be given, and this is used to check
version specifications given in the Imports field when a namespace is loaded.
setClass() has a new argument, slots, clearer and less ambiguous than
representation. It is recommended for future code, but should be back-compatible.
At the same time, the allowed slot specification is slightly more general. See the
documentation for details.
mget() now has a default for envir (the frame from which it is called), for consistency
with get() and assign().
close() now returns an integer status where available, invisibly. (Wish of
The internal method of tar() can now store paths too long for the ustar format,
using the (widely supported) GNU extension. It can also store long link names, but
these are much less widely supported. There is support for larger files, up to the
ustar limit of 8GB.

Local reference classes have been added to package methods. These are a technique
for avoiding unneeded copying of large components of objects while retaining standard
R functional behavior. See ?LocalReferenceClasses.
untar() has a new argument restore_times which if false (not the default) discards
the times in the tarball. This is useful if they are incorrect (some tarballs submitted
to cran have times in a local time zone or many years in the past even though the
standard required them to be in UTC).
replayplot() cannot (and will not attempt to) replay plots recorded under R <
3.0.0. It may crash the R session if an attempt is made to replay plots created in a
different build of R >= 3.0.0.
Palette changes get recorded on the display list, so replaying plots (including when
resizing screen devices and using dev.copy()) will work better when the palette is
changed during a plot.
chol(pivot = TRUE) now defaults to LAPACK, not LINPACK.
The parse() function has a new parameter keep.source, which defaults to
Profiling via Rprof() now optionally records information at the statement level, not
just the function level.
The Rprof() function now quotes function names in in its output file on Windows,
to be consistent with the quoting in Unix.
Profiling via Rprof() now optionally records information about time spent in GC.
The HTML help page for a package now displays non-vignette documentation files in
a more accessible format.
To support options(stringsAsFactors = FALSE), model.frame(),
model.matrix() and replications() now automatically convert character
vectors to factors without a warning.
The print method for objects of class "table" now detects tables with 0-extents and
prints the results as, e.g., < table of extent 0 x 1 x 2 >. (Wish of PR#15198.)
Deparsing involving calls to anonymous functions has been made closer to reversible
by the addition of extra parentheses.
The function utils::packageName() has been added as a lightweight version of
find.package(lib.loc = NULL) now treats loaded namespaces preferentially in the
same way as attached packages have been for a long time.
In Windows, the Change Directory dialog now defaults to the current working direc-
tory, rather than to the last directory chosen in that dialog.
available.packages() gains a "license/restricts_use" filter which retains only
packages for which installation can proceed solely based on packages which are guar-
anteed not to restrict use.
New check_packages_in_dir() function in package tools for conveniently checking
source packages along with their reverse dependencies.
Rs completion mechanism has been improved to handle help requests (starting with a
question mark). In particular, help prefixes are now supported, as well as quoted help
topics. To support this, completion inside quotes are now handled by R by default
on all platforms.

The memory manager now allows the strategy used to balance garbage collection and
memory growth to be controlled by setting the environment variable R_GC_MEM_GROW.
See ?Memory for more details.
(For experts only, as the introductory manual says.) The use of environment vari-
ables R_NSIZE and R_VSIZE to control the initial (= minimum) garbage collection
trigger for number of cons cels and size of heap has been restored: they can be over-
ridden by the command-line options --min-nsize and --min-vsize; see ?Memory.
On Windows, the device name for bitmap devices as reported by .Device and
.Devices no longer includes the file name. This is for consistency with other plat-
forms and was requested by the lattice maintainer.
win.metafile() still uses the file name: the exact form is used by package tkrplot.
set.seed(NULL) re-initializes .Random.seed as done at the beginning of the session
if not already set. (Suggestion of Bill Dunlap.)
The breaks argument in hist.default() can now be a function that returns the
breakpoints to be used (previously it could only return the suggested number of
File share/licenses/licenses.db has some clarifications, especially as to which
variants of BSD and MIT is intended and how to apply them to packages. The
problematic licence Artistic-1.0 has been removed.

LONG VECTORS: This section applies only to 64-bit platforms.

There is support for vectors longer than 231 1 elements. This applies to raw, logical,
integer, double, complex and character vectors, as well as lists. (Elements of character
vectors remain limited to 231 1 bytes.)
Most operations which can sensibly be done with long vectors work: others may
return the error long vectors not supported yet. Most of these are because they
explicitly work with integer indices (e.g., anyDuplicated() and match()) or because
other limits (e.g., of character strings or matrix dimensions) would be exceeded or
the operations would be extremely slow.
length() returns a double for long vectors, and lengths can be set to 231 or more by
the replacement function with a double value.
Most aspects of indexing are available. Generally double-valued indices can be used
to access elements beyond 231 1.
There is some support for matrices and arrays with each dimension less than 231 but
total number of elements more than that. Only some aspects of matrix algebra work
for such matrices, often taking a very long time. In other cases the underlying Fortran
code has an unstated restriction (as was found for complex svd()).
dist() can produce dissimilarity objects for more than 65536 rows (but for example
hclust() cannot process such objects).
serialize() to a raw vector is unlimited in size (except by resources).
The C-level function R_alloc can now allocate 235 or more bytes.
agrep() and grep() will return double vectors of indices for long vector inputs.
Many calls to .C() have been replaced by .Call() to allow long vectors to be sup-
ported (now or in the future). Regrettably several packages had copied the non-API
.C() calls and so failed.

.C() and .Fortran() do not accept long vector inputs. This is a precaution as it is
very unlikely that existing code will have been written to handle long vectors (and
the R wrappers often assume that length(x) is an integer).
Most of the methods for sort() work for long vectors.
rank(), sort.list() and order() support long vectors (slowly except for radix
sample() can do uniform sampling from a long vector.

More use has been made of R objects representing registered entry points, which is
more efficient as the address is provided by the loader once only when the package is
This has been done for packages base, methods, splines and tcltk: it was already
in place for the other standard packages.
Since these entry points are always accessed by the R entry points they do not need
to be in the load table which can be substantially smaller and hence searched faster.
This does mean that .C / .Fortran / .Call calls copied from earlier versions of R
may no longer work but they were never part of the API.
Many .Call() calls in package base have been migrated to .Internal() calls.
solve() makes fewer copies, especially when b is a vector rather than a matrix.
eigen() makes fewer copies if the input has dimnames.
Most of the linear algebra functions make fewer copies when the input(s) are not
double (e.g., integer or logical).
A foreign function call (.C() etc) in a package without a PACKAGE argument will only
look in the first DLL specified in the NAMESPACE file of the package rather than
searching all loaded DLLs. A few packages needed PACKAGE arguments added.
The @<- operator is now implemented as a primitive, which should reduce some
copying of objects when used. Note that the operator object must now be in package
base: do not try to import it explicitly from package methods.

The transitional support for installing packages without namespaces (required since
R 2.14.0) has been removed. R CMD build will still add a namespace, but a
.First.lib() function will need to be converted.
R CMD INSTALL no longer adds a namespace (so installation will fail), and a
.First.lib() function in a package will be ignored (with an installation warning
for now).
As an exception, packages without a R directory and no NAMESPACE file can still be
Packages can specify in their DESCRIPTION file a line like
Biarch: yes
to be installed on Windows with --force-biarch.
Package vignettes can now be processed by other engines besides Sweave; see Writing
R Extensions and the tools::vignetteEngine help topic for details.
The *.R tangled source code for vignettes is now included in tarballs when
R CMD build is used to produce them. In R 3.0.0, *.R files not in the sources
will be produced at install time, but eventually this will be dropped.

The package type "mac.binary" now looks in a path in the repository with-
out any Mac subtype (which used to be universal or leopard): it looks
in bin/macosx/contrib/3.0 rather than bin/macosx/leopard/contrib/2.15).
This is the type used for the cran binary distribution for OS X as from R 3.0.0.
File etc/Makeconf makes more use of the macros $(CC), $(CXX), $(F77) and $(FC),
so the compiler in use can be changed by setting just these (and if necessary the
corresponding flags and FLIBS) in file ~/.R/Makevars.
This is convenient for those working with binary distributions of R, e.g. on OS X.

R CMD check now gives a warning rather than a note if it finds calls to abort, assert
or exit in compiled code, and has been able to find the .o file in which the calls
Such calls can terminate the R process which loads the package.
The location of the build and check environment files can now be specified by the
environment variables R_BUILD_ENVIRON and R_CHECK_ENVIRON, respectively.
R CMD Sweave gains a --compact option to control possibly reducing the size of the
PDF file it creates when --pdf is given.
R CMD build now omits Eclipses .metadata directories, and R CMD check warns if
it finds them.
R CMD check now does some checks on functions defined within reference classes,
including of .Call() etc calls.
R CMD check --as-cran notes assignments to the global environment, calls to
data() which load into the global environment, and calls to attach().
R CMD build by default uses the internal method of tar() to prepare the tarball. This
is more likely to produce a tarball compatible with R CMD INSTALL and R CMD check:
an external tar program, including options, can be specified via the environment
variable R_BUILD_TAR.
tools::massageExamples() is better protected against packages which re-define base
functions such as cat() and get() and so can cause R CMD check to fail when check-
ing examples.
R CMD javareconf has been enhanced to be more similar to the code used by
There is now a test that a JNI program can be compiled (like configure did) and
only working settings are used.
It makes use of custom settings from configuration recorded in etc/javaconf.
The --no-vignettes argument of R CMD build has been renamed to the more
accurate --no-build-vignettes: its action has always been to (re)build vignettes
and never omitted them.
R CMD check accepts --no-build-vignettes as a preferred synonym for


The ENCODING argument to .C() is defunct. Use iconv() instead.
The .Internal(eval.with.vis) non-API function has been removed.
Support for the converters for use with .C() has been removed, including the oft
misused non-API header R_ext/RConverters.h.

The previously deprecated uses of array() with a 0-length dim argument and
tapply() with a 0-length INDEX list are now errors.
Translation packages are defunct.
Calling rep() or on a pairlist or other non-vector object is now an error.
Several non-API entry points have been transferred to packages (e.g., R_zeroin2) or
replaced by different non-API entry points (e.g., R_tabulate).
The internal graphics device invoked by .Call("R_GD_nullDevice", package = "grDevices")
has been removed: use pdf(file = NULL) instead.
The .Fortran() entry point "dqrls" which has not been used by R since version
2.15.1 is no longer available.
Functions traceOn() and traceOff() in package methods are now defunct.
Function CRAN.packages() is finally defunct.
Use of col2rgb(0) is defunct: use par("bg") or NA instead.
The long-defunct functions Rd_parse(), anovalist.lm(), categpry(),
clearNames(), gammaCody(),,, lm.wfit.null(),
manglePackageNames(), mauchley.test(), package.contents(),
print.coefmat(), reshapeLong(), reshapeWide(), tkclose(), tkcmd(),
tkfile.dir(), tkfile.tail(), tkopen(), tkputs(), tkread(), trySilent()
and zip.file.extract() have been removed entirely (but are still documented in
the help system).
The unused dataPath argument to attachNamespace() has been removed.
grid.prompt() has been removed: use devAskNewPage() instead.
The long-deprecated intensities component is no longer returned by hist().
mean() for data frames and sd() for data frames and matrices are defunct.
chol(pivot = FALSE, LINPACK = TRUE), ch2inv(LINPACK = TRUE),
eigen(EISPACK = TRUE), solve(LINPACK = TRUE) and svd(LINPACK = TRUE)
are defunct: LAPACK will be used, with a warning.
The keep.source argument to library() and require() is defunct. This option
needs to be set at install time.
Documentation for real(), as.real() and is.real() has been moved to defunct
and the functions removed.
The maxRasters argument of pdf() (unused since R 2.14.0) has been removed.
The unused fontsmooth argument has been removed from the quartz() device.
All the (non-API) EISPACK entry points in R have been removed.
chol(pivot = TRUE, LINPACK = TRUE) is deprecated.
The long-deprecated use of \synopsis in the Usage section of .Rd files will be
removed in R 3.1.0.
.find.package() and .path.package() are deprecated: only the public versions
without the dot have ever been in the API.
In a packages DESCRIPTION file,
License: X11
is deprecated, since it includes Copyright (C) 1996 X Consortium which cannot be
appropriate for a current R package. Use MIT or BSD 2 clause instead.


The C code underlying base graphics has been migrated to the graphics package (and
hence no longer uses .Internal() calls).
Most of the .Internal() calls used in the stats package have been migrated to C
code in that package.
This means that a number of .Internal() calls which have been used by packages
no longer exist, including .Internal(cor) .Internal(cov), .Internal(optimhess)
and .Internal(update.formula).
Some .External() calls to the base package (really to the R executable or shared
library) have been moved to more appropriate packages. Packages should not have
been using such calls, but some did (mainly those used by integrate()).

PACKAGE parallel:
There is a new function mcaffinity() which allows getting or setting the CPU affinity
mask for the current R process on systems that supports this (currently only Linux
has been tested successfully). It has no effect on systems which do not support process
affinity. Users are not expected to use this function directly (with the exception of
fixing libraries that break affinity settings like OpenBLAS) the function is rather
intended to support affinity control in high-level parallel functions. In the future,
R may supplement lack of affinity control in the OS by its own bookkeeping via
mcaffinity() related to processes and threads it spawns.
mcparallel() has a new argument mc.affinity which attempts to set the affinity
of the child process according to the specification contained therein.
The port used by socket clusters is chosen randomly: this should help to avoid
clashes observed when two users of a multi-user machine try to create a cluster
at the same time. To reproduce the previous behaviour set environment variable

There has been some minor re-organization of the non-API header files. In particular,
Rinternals.h no longer includes the non-API header R_exts/PrtUtil.h, and that
no longer includes R_exts/Print.h.
Passing NULL to .C() is now an error.
.C() and .Fortran() now warn if "single" arguments are used with DUP = FALSE,
as changes to such arguments are not returned to the caller.
C entry points R_qsort and R_qsort_I now have start and end as size_t to allow
them to work with longer vectors on 64-bit platforms. Code using them should be
A few recently added C entry points were missing the remapping to Rf_, notably
Some of the interface pointers formerly available only to are now available to
front-ends on all Unix-alikes: one has been added for the interface to View().
PACKAGE = "" is now an error in .C() etc calls: it was always contrary to the docu-
Entry point rcont2 has been migrated to package stats and so is no longer available.
R_SVN_REVISION in Rversion.h is now an integer (rather than a string) and hence
usable as e.g. #if R_SVN_REVISION < 70000.

The entry points rgb2hsv and hsv2rgb have been migrated to package grDevices and
so are no longer available.
R_GE_version has been increased to 10 and name2col removed (use R_GE_str2col
instead). R internal colour codes are now defined using the typedef rcolor.
The REPROTECT macro now checks that the protect index is valid.
Several non-API entry points no longer used by R have been removed, including
the Fortran entry points chol, chol2inv, cg, ch and rg, and the C entry points
Brent_fmin, fft_factor and fft_work.
If a .External call is registered with a number of arguments (other than -1), the
number of arguments passed is checked for each call (as for other foreign function
It is now possible to write custom connection implementations outside core R using
R_ext/Connections.h. Please note that the implementation of connections is still
considered internal and may change in the future (see the above file for details).

The management of translations has been converted to R code: see
The translations for the R interpreter and RGui.exe are now part of the base package
(rather than having sources in directory po and being installed to share/locale).
Thus the base package supports three translation domains, R-base, R and RGui.
The compiled translations which ship with R are all installed to the new package
translations for easier updating. The first package of that name found on .libPaths()
at the start of the R session will be used. (It is possible messages will be used
before .libPaths() is set up in which case the default translations will be used:
set environment variable R_TRANSLATIONS to point to the location of the intended
translations package to use this right from the start.)
The translations form a separate group in the Windows installer, so can be omitted
if desired.
The markup for many messages has been changed to make them easier to translate,
incorporating suggestions from Lukasz Daniel.

There is again support for building without using the C long double type. This is
required by C99, but system implementations can be slow or flawed. Use configure
option --disable-long-double.
make pdf and make install-pdf now make and install the full reference index (in-
cluding all base and recommended packages).
The reference manual on the Windows GUI menu and included in the installer is
now the full reference index, including all base and recommended packages.
R help pages and manuals have no ISBNs because ISBN rules no longer allow con-
stantly changing content to be assigned an ISBN.
The Windows installer no longer installs a Start Menu link to the static help pages;
as most pages are generated dynamically, this led to a lot of broken links.
Any custom settings for Java configuration are recorded in file etc/javaconf for
subsequent use by R CMD javareconf.

There is now support for makeinfo version 5.0 (which requires a slightly different
.texi syntax).
The minimum versions for --use-system-zlib and --use-system-pcre are now
tested as 1.2.5 and 8.10 respectively.
On Windows, the stack size is reduced to 16MB on 32-bit systems: misguided users
were launching many threads without controlling the stack size.
configure no longer looks for file ~/.Rconfig: ~/.R/config has long been pre-

When R CMD build is run in an encoding other than the one specified in the pack-
ages DESCRIPTION file it tries harder to expand the authors@R field in the specified
encoding. (PR#14958)
If R CMD INSTALL is required to expand the authors@R field of the DESCRIPTION
file, it tries harder to do so in the encoding specified for the package (rather than
using ASCII escapes).
Fix in package grid for pushing a viewport into a layout cell, where the layout is
within a viewport that has zero physical width OR where the layout has zero total
relative width (likewise for height). The layout column widths (or row heights) in this
case were being calculated with non-finite values. (Reported by Winston Chang.)
solve(A, b) for a vector b gave the answer names from colnames(A) for
LINPACK = TRUE but not in the default case.
La.svd() accepts logical matrices (as documented, and as svd() did).
legend() now accepts negative pch values, in the same way points() long has.
Parse errors when installing files now correctly display the name of the file containing
the bad code.
In Windows, tcltk windows were not always properly constructed. (PR#15150)
The internal functions implementing parse(), tools::parseLatex() and
tools::parse_Rd() were not reentrant, leading to errors in rare circumstances such
as a garbage collection triggering a recursive call.
Field assignments in reference class objects via $<- were not being checked because
the magic incantation to turn methods on for that primitive operator had been inad-
vertently omitted.
setHook(hookname, value, action="replace") set the hook to be the value,
rather than a list containing the value as documented. (PR#15167)
If a package used a NEWS.Rd file, the main HTML package index page did not link
to it. (Reported by Dirk Eddelbuettel.)
The primitive implementation of @<- was not checking the class of the replacement.
It now does a check, quicker but less general than slot<-. See the help.
split(x, f) now recycles classed objects x in the same way as vectors. (Reported
by Martin Morgan.)
pbeta(.28, 1/2, 2200, lower.tail=FALSE, log.p=TRUE) is no
longer -Inf; ditto for corresponding pt() and pf() calls, such as
pt(45, df=5000, lower.tail=FALSE,log.p=TRUE). (PR#15162)
The Windows graphics device would crash R if a user attempted to load the graphics
history from a variable that was not a saved history. (PR#15230)

The workspace size for the predict() method for loess() could exceed the maximum
integer size. (Reported by Hiroyuki Kawakatsu.)
ftable(x, row.vars, col.vars) now also works when the *.vars arguments are
(integer or character vectors) of length zero.
Calling cat() on a malformed UTF-8 string could cause the Windows GUI to lock
up. (PR#15227)
removeClass(cc) gave node stack overflow for some class definitions containing
"array" or "matrix".

CHANGES in previous versions

Older news can be found in text format in files NEWS.0, NEWS.1 and NEWS.2 in
the doc directory. News in HTML format for R versions from 2.10.0 to 2.15.3 is in